diff --git "a/csv/claims.csv" "b/csv/claims.csv"
new file mode 100644--- /dev/null
+++ "b/csv/claims.csv"
@@ -0,0 +1,532 @@
+id,vendor,source_key,statement,verbatim,compared,stated,in_file,outcome,note,label_mismatch,source_title,source_url,source_fetched,source_round
+C1,NVIDIA,NVIDIA v5.1,GB300 NVL72 delivered 45% higher performance per GPU on the new DeepSeek-R1 benchmark in the offline scenario,true,"entries 5.1-0072 / 5.1-0071, tokens/s divided by 72",+45%,+45.2% (5842.5 vs 4023.8 tokens/s/GPU),CONSISTENT,rounds to the stated figure,false,NVIDIA Technical Blog: NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut,https://developer.nvidia.com/blog/nvidia-blackwell-ultra-sets-new-inference-records-in-mlperf-debut,2026-10-28,MLPerf Inference v5.1
+C2,NVIDIA,NVIDIA v5.1,and 25% in the server scenario,true,"entries 5.1-0072 / 5.1-0071, tokens/s divided by 72",+25%,+24.9% (2907.3 vs 2327.5 tokens/s/GPU),CONSISTENT,rounds to the stated figure,false,NVIDIA Technical Blog: NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut,https://developer.nvidia.com/blog/nvidia-blackwell-ultra-sets-new-inference-records-in-mlperf-debut,2026-10-28,MLPerf Inference v5.1
+C3,NVIDIA,NVIDIA v5.1,"Table 2 in the post lists per-GPU DeepSeek-R1 throughput in tokens/s: GB200 NVL72 4,024 offline and 2,327 server, GB300 NVL72 5,842 offline and 2,907 server.",false,four cells recomputed from the summary file,"4024, 2327, 5842, 2907","4023.78, 2327.47, 5842.49, 2907.33",CONSISTENT,all four cells match after rounding,false,NVIDIA Technical Blog: NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut,https://developer.nvidia.com/blog/nvidia-blackwell-ultra-sets-new-inference-records-in-mlperf-debut,2026-10-28,MLPerf Inference v5.1
+C6,NVIDIA,NVIDIA v5.1,a greater than 5x cumulative improvement compared to in-flight batching results collected on a DGX H200 system,true,"entries 5.1-0071 / 5.1-0075 (8xH200, Nebius, closed), Llama 3.1 405B Interactive",greater than 5x,5.44x (137.79 vs 25.35 tokens/s/GPU),CONSISTENT,Hopper row is a verified closed row. Precision differs: weights fp4 against fp8. The blog names the Hopper system DGX H200 and the file lists the entry as an 8xH200 HGX submission by Nebius.,false,NVIDIA Technical Blog: NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut,https://developer.nvidia.com/blog/nvidia-blackwell-ultra-sets-new-inference-records-in-mlperf-debut,2026-10-28,MLPerf Inference v5.1
+C7I,NVIDIA,NVIDIA v5.1,"Table 1 in the post lists a per-GPU record of 138 tokens/s for Llama 3.1 405B, interactive.",false,maximum over all closed NVIDIA-accelerator rows for Llama 3.1 405B Interactive in v5.1,138,"137.79 (5.1-0071, NVIDIA, NVIDIA GB200)",CONSISTENT,the stated figure is the highest per-GPU value in the closed v5.1 rows,false,NVIDIA Technical Blog: NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut,https://developer.nvidia.com/blog/nvidia-blackwell-ultra-sets-new-inference-records-in-mlperf-debut,2026-10-28,MLPerf Inference v5.1
+C7O,NVIDIA,NVIDIA v5.1,"Table 1 in the post lists a per-GPU record of 224 tokens/s for Llama 3.1 405B, offline.",false,maximum over all closed NVIDIA-accelerator rows for Llama 3.1 405B Offline in v5.1,224,"223.67 (5.1-0072, NVIDIA, NVIDIA GB300)",CONSISTENT,the stated figure is the highest per-GPU value in the closed v5.1 rows,false,NVIDIA Technical Blog: NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut,https://developer.nvidia.com/blog/nvidia-blackwell-ultra-sets-new-inference-records-in-mlperf-debut,2026-10-28,MLPerf Inference v5.1
+C7S,NVIDIA,NVIDIA v5.1,"Table 1 in the post lists a per-GPU record of 170 tokens/s for Llama 3.1 405B, server.",false,maximum over all closed NVIDIA-accelerator rows for Llama 3.1 405B Server in v5.1,170,"170.12 (5.1-0072, NVIDIA, NVIDIA GB300)",CONSISTENT,the stated figure is the highest per-GPU value in the closed v5.1 rows,false,NVIDIA Technical Blog: NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut,https://developer.nvidia.com/blog/nvidia-blackwell-ultra-sets-new-inference-records-in-mlperf-debut,2026-10-28,MLPerf Inference v5.1
+C4,NVIDIA,NVIDIA v5.1,"compared to unverified results collected on a Hopper-based system, Blackwell Ultra delivered about 5x higher throughput per GPU",true,"Table 2 Hopper cells (1,253 / 556) vs GB300, and vs the only H200 DeepSeek-R1 row in the file",about 5x,"4.66x offline and 5.23x server against NVIDIA's Hopper cells. 5.25x offline against the one published H200 row (5.1-0375, HPE, open, 1113.1 tokens/s/GPU)",CANNOT_TELL,NVIDIA states the Hopper cells are not verified by MLCommons and no closed-division H200 DeepSeek-R1 row exists in the file. Against the published open-division row the ratio is about 5x.,false,NVIDIA Technical Blog: NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut,https://developer.nvidia.com/blog/nvidia-blackwell-ultra-sets-new-inference-records-in-mlperf-debut,2026-10-28,MLPerf Inference v5.1
+C5,NVIDIA,NVIDIA v5.1,nearly 1.5x increase in throughput per GPU compared to traditional aggregated serving using in-flight batching on a DGX B200 system,true,"entries 5.1-0071 (72 GPUs) / 5.1-0069 (8 GPUs), Llama 3.1 405B Interactive",nearly 1.5x,"1.467x (137.79 vs 93.92 tokens/s/GPU). Same two systems, Server: 1.040x, Offline: 1.011x",CANNOT_TELL,"The ratio holds. The two entries differ in system size and rack, so the file cannot separate disaggregated serving from the 72-GPU system. Whether entry 5.1-0071 used disaggregated serving is not stated in the summary file or the entry README.",false,NVIDIA Technical Blog: NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut,https://developer.nvidia.com/blog/nvidia-blackwell-ultra-sets-new-inference-records-in-mlperf-debut,2026-10-28,MLPerf Inference v5.1
+NT6,NVIDIA,NVIDIA Training v5.0,"the TP4 – CP2 – PP8 – DP39 layout emerges as optimal for the NVIDIA 2,496-GPU-scale Blackwell submission",true,Blackwell Llama 3.1 405B entries above 512 GPUs in the v5.0 training times,"a 2,496 GPU Blackwell submission by NVIDIA",IBM+CoreWeave+NVIDIA 1024 GPUs 62.07 min. IBM+CoreWeave+NVIDIA 1536 GPUs 42.46 min. IBM+CoreWeave+NVIDIA 2048 GPUs 32.63 min. IBM+CoreWeave+NVIDIA 2496 GPUs 27.34 min,BELOW,"The 2,496 GPU Blackwell entry is submitted by IBM, CoreWeave and NVIDIA together. NVIDIA alone has no Blackwell entry above 512 GPUs. 64 GPUs per rack times 39 data parallel ranks is 2,496.",false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers up to 2.6x Higher Performance in MLPerf Training v5.0 (post dated 4 June 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-up-to-2-6x-higher-performance-in-mlperf-training-v5-0/,2026-10-28,MLPerf Training v5.0
+NT1,NVIDIA,NVIDIA Training v5.0,"Table 1 in the post lists time to train in minutes on the NVIDIA platform: Llama 3.1 405B 20.8, Llama 2 70B LoRA 0.56, Stable Diffusion 1.04, R-GAT 0.84, DLRM-DCNv2 0.7, BERT 0.3, RetinaNet 1.4.",false,"training_v50_times.json, entries at the scale of the cited submissions, official-style time","20.8, 0.56, 1.04, 0.84, 0.7, 0.3 and 1.4 minutes",llama31_405b on 8192 GPUs 20.81. llama2_70b_lora on 512 GPUs 0.56. stable_diffusion on 512 GPUs 1.04. rgat on 256 GPUs 0.84. dlrm_dcnv2 on 64 GPUs 0.71. bert on 512 GPUs 0.30. retinanet on 512 GPUs 1.39,CONSISTENT,"The 20.8 minute Llama 3.1 405B value is the 8,192 GPU H100 entry. It is not a Blackwell result. The DLRM value is the Supermicro 64 GPU entry. Times are olympic averages of the logged runs, recomputed.",false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers up to 2.6x Higher Performance in MLPerf Training v5.0 (post dated 4 June 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-up-to-2-6x-higher-performance-in-mlperf-training-v5-0/,2026-10-28,MLPerf Training v5.0
+NT2,NVIDIA,NVIDIA Training v5.0,"Table 2 in the post lists Llama 3.1 405B on 512 GPUs: Hopper 269.12 minutes, Blackwell 121.09 minutes, speedup 2.2x.",false,"NVIDIA 512 GPU Llama 3.1 405B entries, H100 and Blackwell","269.12 and 121.09 minutes, 2.2x","H100 269.12, Blackwell 121.09 (submitted by IBM+CoreWeave+NVIDIA), ratio 2.22x. The NVIDIA own 512 GPU Blackwell entry is 121.76",CONSISTENT,"The 121.09 minute Blackwell value is the fastest 512 GPU Blackwell entry. NVIDIA own 512 GPU entry is 121.76, which gives 2.21x.",false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers up to 2.6x Higher Performance in MLPerf Training v5.0 (post dated 4 June 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-up-to-2-6x-higher-performance-in-mlperf-training-v5-0/,2026-10-28,MLPerf Training v5.0
+NT3,NVIDIA,NVIDIA Training v5.0,"Tables 3, 5 and 6 in the post list eight-GPU Blackwell times of 11.14 minutes (Llama 2 70B LoRA), 12.86 minutes (Stable Diffusion v2) and 4.97 minutes (RetinaNet), with speedups over Hopper of 2.51x, 2.64x and 2.25x.",false,NVIDIA eight-GPU Blackwell entries in the v5.0 training times. Hopper values are v4.1 entries,three Blackwell times and speedups,"Llama 2 70B LoRA: Blackwell 11.14 (post 11.14), speedup on the post Hopper time 2.51x. Stable Diffusion v2: Blackwell 12.86 (post 12.86), speedup on the post Hopper time 2.64x. GNN R-GAT: Blackwell 4.97 (post 4.97), speedup on the post Hopper time 2.25x",CONSISTENT,"The Hopper times are v4.1 entries that are not in the files, so the speedups rest on the post values for those. The Blackwell times match.",false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers up to 2.6x Higher Performance in MLPerf Training v5.0 (post dated 4 June 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-up-to-2-6x-higher-performance-in-mlperf-training-v5-0/,2026-10-28,MLPerf Training v5.0
+NT5,NVIDIA,NVIDIA Training v5.0,"In this latest round of MLPerf Training, the NVIDIA platform delivered the fastest time to train across all seven benchmarks.",true,"fastest time per benchmark in the v5.0 training times, any submitter, NVIDIA accelerators counted as the NVIDIA platform",fastest on all seven,"bert: NVIDIA 512 GPUs 0.3 min, dlrm_dcnv2: Supermicro 64 GPUs 0.71 min, llama2_70b_lora: NVIDIA 512 GPUs 0.56 min, llama31_405b: NVIDIA 8,192 GPUs 20.81 min, retinanet: NVIDIA 512 GPUs 1.39 min, rgat: NVIDIA 256 GPUs 0.84 min, stable_diffusion: NVIDIA 512 GPUs 1.04 min",CONSISTENT,Fastest is read as the shortest time at any GPU count. Every benchmark has its fastest entry on NVIDIA accelerators.,false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers up to 2.6x Higher Performance in MLPerf Training v5.0 (post dated 4 June 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-up-to-2-6x-higher-performance-in-mlperf-training-v5-0/,2026-10-28,MLPerf Training v5.0
+NT4,NVIDIA,NVIDIA Training v5.0,Table 4 in the post lists June 2025 unverified results of 21.84 minutes on H200 and 10.34 minutes on Blackwell for Llama 2 70B LoRA.,false,unverified results are not entries in the files,21.84 and 10.34 minutes unverified,no entry,CANNOT_TELL,The post marks the results as not verified by MLCommons.,false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers up to 2.6x Higher Performance in MLPerf Training v5.0 (post dated 4 June 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-up-to-2-6x-higher-performance-in-mlperf-training-v5-0/,2026-10-28,MLPerf Training v5.0
+NW1,NVIDIA,NVIDIA Training v5.1,"Table 1 in the post lists time to train and maximum submission scale: Llama 3.1 405B 10 minutes on 5,120 Blackwell GPUs, Llama 3.1 8B 5.2 minutes on 512 Blackwell Ultra GPUs, Llama 2 70B LoRA 0.40 minutes on 512 Blackwell Ultra GPUs, FLUX.1 12.5 minutes on 1,152 Blackwell GPUs, DLRM-DCNv2 0.71 minutes on 64 Blackwell GPUs, R-GAT 0.84 minutes on 256 Blackwell GPUs, RetinaNet 1.4 minutes on 512 Blackwell GPUs.",false,"training_v51_times.json, fastest entry per benchmark on NVIDIA accelerators (DLRM-DCNv2 from the v5.0 file), official-style time and GPU count","10, 5.2, 0.40, 12.5, 0.71, 0.84 and 1.4 minutes at 5,120, 512, 512, 1,152, 64, 256 and 512 GPUs",llama31_405b NVIDIA 5120 GPUs 9.995. llama31_8b NVIDIA 512 GPUs 5.219. llama2_70b_lora NVIDIA 512 GPUs 0.403. flux1 NVIDIA 1152 GPUs 12.514. dlrm_dcnv2 Supermicro 64 GPUs 0.713. rgat NVIDIA 256 GPUs 0.837. retinanet NVIDIA 512 GPUs 1.388,CONSISTENT,Mismatches: none. The scale is the GPU count of the fastest entry. The DLRM-DCNv2 row is the v5.0 entry (the post footnote lists 5.0-0082) and the fastest v5.1 DLRM entry is 0.77 min.,false,NVIDIA Technical Blog: NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks (post dated 12 November 2025),https://developer.nvidia.com/blog/nvidia-blackwell-architecture-sweeps-mlperf-training-v5-1-benchmarks/,2026-10-28,MLPerf Training v5.1
+NW2,NVIDIA,NVIDIA Training v5.1,The NVIDIA Blackwell architecture powered the fastest time to train across every MLPerf Training v5.1 benchmark,true,"fastest entry per benchmark in training_v51_times.json, any submitter, NVIDIA accelerators counted as the NVIDIA platform",fastest time to train on every benchmark,"llama31_405b: NVIDIA on NVIDIA accelerators, llama31_8b: NVIDIA on NVIDIA accelerators, llama2_70b_lora: NVIDIA on NVIDIA accelerators, flux1: NVIDIA on NVIDIA accelerators, dlrm_dcnv2: NVIDIA on NVIDIA accelerators, rgat: NVIDIA on NVIDIA accelerators, retinanet: NVIDIA on NVIDIA accelerators",CONSISTENT,The post says at maximum scale and at each submitted scale. Only the maximum scale (fastest overall) is recomputed here.,false,NVIDIA Technical Blog: NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks (post dated 12 November 2025),https://developer.nvidia.com/blog/nvidia-blackwell-architecture-sweeps-mlperf-training-v5-1-benchmarks/,2026-10-28,MLPerf Training v5.1
+NW3,NVIDIA,NVIDIA Training v5.1,NVIDIA was also the only platform to submit to this benchmark,true,accelerators of all FLUX.1 entries in training_v51_times.json,only the NVIDIA platform submitted to FLUX.1,NVIDIA on NVIDIA Blackwell GPU (GB200). NVIDIA on NVIDIA Blackwell Ultra GPU (GB300). Nebius on NVIDIA Blackwell GPU (B200-SXM-180GB). Oracle on NVIDIA Blackwell GPU (B200-SXM-180GB). Supermicro on NVIDIA Blackwell Ultra GPU (GB300). University-of-Florida on NVIDIA Blackwell GPU (B200-SXM-180GB),CONSISTENT,"Other submitters appear, all on NVIDIA accelerators. Platform is read as accelerator.",false,NVIDIA Technical Blog: NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks (post dated 12 November 2025),https://developer.nvidia.com/blog/nvidia-blackwell-architecture-sweeps-mlperf-training-v5-1-benchmarks/,2026-10-28,MLPerf Training v5.1
+NW4,NVIDIA,NVIDIA Training v5.1,"NVIDIA set a new time-to-train record of 10 minutes powered by 5,120 Blackwell GPUs. This is a 2.7x increase compared to the fastest submission using Blackwell GPUs last round",true,"fastest Blackwell Llama 3.1 405B entry in training_v50_times.json over the v5.1 5,120 GPU entry","10 minutes, 2.7x",v5.1 9.995 min on 5120 GPUs. v5.0 fastest Blackwell 27.34 min on 2496 GPUs (IBM+CoreWeave+NVIDIA). Ratio 2.735x,CONSISTENT,"The v5.0 entry was submitted by IBM, CoreWeave and NVIDIA together.",false,NVIDIA Technical Blog: NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks (post dated 12 November 2025),https://developer.nvidia.com/blog/nvidia-blackwell-architecture-sweeps-mlperf-training-v5-1-benchmarks/,2026-10-28,MLPerf Training v5.1
+NW5,NVIDIA,NVIDIA Training v5.1,"the submission using 2,560 Blackwell GPUs achieved a score of 18.79 minutes. This is 3x faster than the previous NVIDIA submissions with the same number of NVIDIA Hopper architecture GPUs",true,"2,560 GPU Blackwell entry in training_v51_times.json and 2,560 GPU H100 entry in training_v50_times.json","18.79 minutes, 3x",Blackwell 18.794. H100 57.54. Ratio 3.06x,CONSISTENT,3.06x rounds to 3x.,false,NVIDIA Technical Blog: NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks (post dated 12 November 2025),https://developer.nvidia.com/blog/nvidia-blackwell-architecture-sweeps-mlperf-training-v5-1-benchmarks/,2026-10-28,MLPerf Training v5.1
+NW6,NVIDIA,NVIDIA Training v5.1,"Effective performance per Blackwell GPU also increased by 42%, when comparing the performance of the 2,496 Blackwell GPU submission last round to the 2,560 Blackwell GPU submission this round",true,"27.34 over 18.79 times 2,496 over 2,560, from the two entries",42%,"27.335 min on 2,496 GPUs and 18.794 min on 2,560 GPUs give 41.8%",CONSISTENT,The post says performance per GPU is derived by dividing the ratios of delivered performance and scales submitted.,false,NVIDIA Technical Blog: NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks (post dated 12 November 2025),https://developer.nvidia.com/blog/nvidia-blackwell-architecture-sweeps-mlperf-training-v5-1-benchmarks/,2026-10-28,MLPerf Training v5.1
+NW7,NVIDIA,NVIDIA Training v5.1,"Scaling efficiency, which is the amount of performance increase with additional GPUs, scaled by 10x from 512 Blackwell GPUs to 5,120 Blackwell GPUs, was 85%",true,"512 GPU GB200 entry over 5,120 GPU entry, divided by 10",85%,"512 GPUs 84.94 min, 5,120 GPUs 9.995 min, efficiency 85.0%",CONSISTENT,The 512 GPU Blackwell Ultra entry (64.6 min) is a different accelerator and is not used.,false,NVIDIA Technical Blog: NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks (post dated 12 November 2025),https://developer.nvidia.com/blog/nvidia-blackwell-architecture-sweeps-mlperf-training-v5-1-benchmarks/,2026-10-28,MLPerf Training v5.1
+NW8,NVIDIA,NVIDIA Training v5.1,NVIDIA makes the industry’s first FP4 training submissions with NVFP4,true,no earlier training rounds record numeric formats in the files,first FP4 training submissions,the times files do not hold numeric formats,CANNOT_TELL,"Numeric format is in the submission code, which is not in the files used here.",false,NVIDIA Technical Blog: NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks (post dated 12 November 2025),https://developer.nvidia.com/blog/nvidia-blackwell-architecture-sweeps-mlperf-training-v5-1-benchmarks/,2026-10-28,MLPerf Training v5.1
+NW9,NVIDIA,NVIDIA Training v5.1,Figure 1 in the post shows relative Llama 3.1 405B pretraining and Llama 2 70B LoRA fine-tuning performance at 512-GPU and 8-GPU scales.,false,figure values are not printed in the post,figure 1 relative performance,figure values not in the text,CANNOT_TELL,"The post cites entries 4.1-0050, 5.0-0076, 5.0-0067, 5.1-0058 and 5.1-0060 and prints no numbers for the figure.",false,NVIDIA Technical Blog: NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks (post dated 12 November 2025),https://developer.nvidia.com/blog/nvidia-blackwell-architecture-sweeps-mlperf-training-v5-1-benchmarks/,2026-10-28,MLPerf Training v5.1
+NX1,NVIDIA,NVIDIA Training v5.1 per dollar,GB200 NVL72 delivered up to 3.2x faster training performance on the Llama 3.1 405B benchmark at the same GPU count compared to optimized submissions using NVIDIA Hopper running FP8,true,512 GPU H100 entry in training_v50_times.json over the 512 GPU GB200 entry in training_v51_times.json,up to 3.2x,"H100 269.12 min, GB200 84.94 min, ratio 3.17x",CONSISTENT,The H100 entry is from v5.0. The files hold no later NVIDIA Hopper 512 GPU entry.,false,NVIDIA Technical Blog: NVIDIA Blackwell Enables 3x Faster Training and Nearly 2x Training Performance Per Dollar than Previous-Gen Architecture (post dated 11 December 2025),https://developer.nvidia.com/blog/nvidia-blackwell-enables-3x-faster-training-and-nearly-2x-training-performance-per-dollar-than-previous-gen-architecture/,2026-10-28,MLPerf Training v5.1
+NX2,NVIDIA,NVIDIA Training v5.1 per dollar,up to 1.4x higher performance at similar scale,true,512 GPU Blackwell entries in training_v50_times.json over the 512 GPU GB200 entry in training_v51_times.json,up to 1.4x,"v5.0 NVIDIA 121.76 min and fastest v5.0 121.09 min, v5.1 GB200 84.94 min, ratios 1.43x and 1.43x",CONSISTENT,Both v5.0 512 GPU Blackwell entries give 1.4x after rounding.,false,NVIDIA Technical Blog: NVIDIA Blackwell Enables 3x Faster Training and Nearly 2x Training Performance Per Dollar than Previous-Gen Architecture (post dated 11 December 2025),https://developer.nvidia.com/blog/nvidia-blackwell-enables-3x-faster-training-and-nearly-2x-training-performance-per-dollar-than-previous-gen-architecture/,2026-10-28,MLPerf Training v5.1
+NX3,NVIDIA,NVIDIA Training v5.1 per dollar,"GB300 NVL72 completed the Llama 3.1 405B benchmark 1.9x faster than GB200 NVL72 at the same scale did last round, bringing the cumulative performance gain compared to Hopper up to 4.2x",true,512 GPU GB300 entry in training_v51_times.json against the 512 GPU Blackwell entries and the 512 GPU H100 entry in training_v50_times.json,1.9x and 4.2x,GB300 64.61 min. GB200 last round 121.76 (ratio 1.88x) and 121.09 (ratio 1.87x). H100 269.12 (ratio 4.17x),CONSISTENT,Both last-round Blackwell 512 GPU entries give 1.9x after rounding.,false,NVIDIA Technical Blog: NVIDIA Blackwell Enables 3x Faster Training and Nearly 2x Training Performance Per Dollar than Previous-Gen Architecture (post dated 11 December 2025),https://developer.nvidia.com/blog/nvidia-blackwell-enables-3x-faster-training-and-nearly-2x-training-performance-per-dollar-than-previous-gen-architecture/,2026-10-28,MLPerf Training v5.1
+NX4,NVIDIA,NVIDIA Training v5.1 per dollar,GB200 NVL72 delivers almost 2x the performance per dollar of H100,true,"performance per dollar needs rental prices, which are not in the files",almost 2x performance per dollar,prices are not in the files,CANNOT_TELL,The post says performance per dollar is derived from published on-demand instance pricing and is not verified by MLCommons.,false,NVIDIA Technical Blog: NVIDIA Blackwell Enables 3x Faster Training and Nearly 2x Training Performance Per Dollar than Previous-Gen Architecture (post dated 11 December 2025),https://developer.nvidia.com/blog/nvidia-blackwell-enables-3x-faster-training-and-nearly-2x-training-performance-per-dollar-than-previous-gen-architecture/,2026-10-28,MLPerf Training v5.1
+NX5,NVIDIA,NVIDIA Training v5.1 per dollar,NVIDIA made the industry’s first and only submissions using FP4 precision,true,numeric formats are not in the times files,first and only FP4 submissions,numeric formats not in the files,CANNOT_TELL,"Numeric format is in the submission code, which is not in the files used here.",false,NVIDIA Technical Blog: NVIDIA Blackwell Enables 3x Faster Training and Nearly 2x Training Performance Per Dollar than Previous-Gen Architecture (post dated 11 December 2025),https://developer.nvidia.com/blog/nvidia-blackwell-enables-3x-faster-training-and-nearly-2x-training-performance-per-dollar-than-previous-gen-architecture/,2026-10-28,MLPerf Training v5.1
+BV1,Broadcom,Broadcom v5.1,"Table 3 in the post lists MLPerf Inference 5.1 results using eight vGPUs in VCF 9.0 on a Dell PowerEdge XE9680 with eight H200 GPUs: Llama 3.1 405B Server 277 and Offline 547 tokens/s, Llama 2 70B Server 33385 and Offline 34301, Llama 2 70B high accuracy Server 33371 and Offline 34486, SDXL Server 17.95 and Offline 18.64 samples/s.",false,"raw/inf_5.1.json, entry 5.1-0029 (Dell_Broadcom), closed datacenter","277, 547, 33385, 34301, 33371, 34486, 17.95 and 18.64",llama3.1-405b Server 277.33. llama3.1-405b Offline 547.26. llama2-70b-99 Server 33384.70. llama2-70b-99 Offline 34301.90. llama2-70b-99.9 Server 33370.60. llama2-70b-99.9 Offline 34485.70. stable-diffusion-xl Server 17.95. stable-diffusion-xl Offline 18.64,ABOVE,The post prints whole tokens per second for the Llama rows. Mismatches: llama2-70b-99 Offline printed 34301 file 34301.9. A printed value that is the file value cut off instead of rounded differs by less than 1 token per second.,false,Broadcom VMware Cloud Foundation Blog: The New Paradigm: MLPerf Inference 5.1 Confirms VCF is the Future of AI/ML Performance (post dated 15 December 2025),https://blogs.vmware.com/cloud-foundation/2025/12/15/mlperf-5-1-confirms-vcf-future-of-ai-ml-performance/,2026-10-28,MLPerf Inference v5.1
+BV2,Broadcom,Broadcom v5.1,VCF and bare metal are only 1% or 2% different in performance,true,"Dell_Broadcom 5.1-0029 against Dell 5.1-0026 (PowerEdge XE9680, eight H200), Llama 2 70B",1% or 2% different,"llama2-70b-99 Server +0.42%, llama2-70b-99 Offline -2.87%, llama2-70b-99.9 Server +0.38%, llama2-70b-99.9 Offline -2.35%",BELOW,Two of four Llama 2 70B Offline differences round to 3% and 2%. The Offline 99 difference is -2.87%.,false,Broadcom VMware Cloud Foundation Blog: The New Paradigm: MLPerf Inference 5.1 Confirms VCF is the Future of AI/ML Performance (post dated 15 December 2025),https://blogs.vmware.com/cloud-foundation/2025/12/15/mlperf-5-1-confirms-vcf-future-of-ai-ml-performance/,2026-10-28,MLPerf Inference v5.1
+BV3,Broadcom,Broadcom v5.1,Because Dell submitted only Llama 2 70B results,true,models in Dell entry 5.1-0026 and Dell_Broadcom entry 5.1-0029,Dell submitted only Llama 2 70B on this system,"Dell 5.1-0026: llama2-70b-99, llama2-70b-99.9, retinanet. Broadcom 5.1-0029: llama2-70b-99, llama2-70b-99.9, llama3.1-405b, stable-diffusion-xl",BELOW,"Dell 5.1-0026 also holds RetinaNet. Broadcom did not submit RetinaNet on this system, so Llama 2 70B is the only model in common.",false,Broadcom VMware Cloud Foundation Blog: The New Paradigm: MLPerf Inference 5.1 Confirms VCF is the Future of AI/ML Performance (post dated 15 December 2025),https://blogs.vmware.com/cloud-foundation/2025/12/15/mlperf-5-1-confirms-vcf-future-of-ai-ml-performance/,2026-10-28,MLPerf Inference v5.1
+BV4,Broadcom,Broadcom v5.1,Llama 3.1 405B Interactive scenario result is unverified by MLCommons Association. Broadcom & SuperMicro did not submit it as it was not required.,true,Broadcom_Supermicro rows with Llama 3.1 405B Interactive,no Interactive submission,0 Interactive rows for Broadcom_Supermicro,CONSISTENT,Whether it was required is not in the files. The file holds no such row.,false,Broadcom VMware Cloud Foundation Blog: The New Paradigm: MLPerf Inference 5.1 Confirms VCF is the Future of AI/ML Performance (post dated 15 December 2025),https://blogs.vmware.com/cloud-foundation/2025/12/15/mlperf-5-1-confirms-vcf-future-of-ai-ml-performance/,2026-10-28,MLPerf Inference v5.1
+BV5,Broadcom,Broadcom v5.1,Stable Diffusion benchmark results were not submitted by SuperMicro on a bare metal platform,true,Supermicro closed datacenter rows with SDXL on eight B200,no SuperMicro bare metal SDXL on eight B200,0 rows,CONSISTENT,"The comparison the post makes instead is to a different submitter, which Figure 1 shows and the text does not number.",false,Broadcom VMware Cloud Foundation Blog: The New Paradigm: MLPerf Inference 5.1 Confirms VCF is the Future of AI/ML Performance (post dated 15 December 2025),https://blogs.vmware.com/cloud-foundation/2025/12/15/mlperf-5-1-confirms-vcf-future-of-ai-ml-performance/,2026-10-28,MLPerf Inference v5.1
+BV6,Broadcom,Broadcom v5.1,"AI/ML Inference workloads from diverse domains of LLM (Llama 3.1 with 405 billion parameters, Speech2Text (Whisper from OpenAI), Text2Images (Stable Diffusion XL)) get on par bare metal performance in VCF",true,Broadcom_Supermicro 5.1-0009 against Supermicro bare metal 5.1-0087 (SYS-422GA-NBRT-LCC),on par bare metal,"llama3.1-405b Offline +0.55%, llama3.1-405b Server +0.28%, whisper Offline -2.10%",CANNOT_TELL,The post gives no threshold for on par. Figure 1 values are not printed. SDXL has no bare metal pair on the same hardware.,false,Broadcom VMware Cloud Foundation Blog: The New Paradigm: MLPerf Inference 5.1 Confirms VCF is the Future of AI/ML Performance (post dated 15 December 2025),https://blogs.vmware.com/cloud-foundation/2025/12/15/mlperf-5-1-confirms-vcf-future-of-ai-ml-performance/,2026-10-28,MLPerf Inference v5.1
+BV7,Broadcom,Broadcom v5.1,achieving near bare-metal performance while benefiting from VCF’s manageability and flexibility,true,"Intel_VMware 5.1-0059 against Intel 5.1-0058 (1-node-2S-GNR_86C), same processor",near bare-metal,"llama3.1-8b Offline -2.97%, llama3.1-8b Server -10.00%, rgat Offline -7.77%, whisper Offline -4.12%",CANNOT_TELL,The post gives no threshold for near. The differences run from -10.0% to -3.0%.,false,Broadcom VMware Cloud Foundation Blog: The New Paradigm: MLPerf Inference 5.1 Confirms VCF is the Future of AI/ML Performance (post dated 15 December 2025),https://blogs.vmware.com/cloud-foundation/2025/12/15/mlperf-5-1-confirms-vcf-future-of-ai-ml-performance/,2026-10-28,MLPerf Inference v5.1
+BV8,Broadcom,Broadcom v5.1,SuperMicro GPU SuperServer AS-4126GS-NBR-LCC with eight NVIDIA NVLinked NVIDIA HGX B200 GPUs in DirectPath I/O,true,System and processor fields of entry 5.1-0009,virtual system is AS-4126GS-NBR-LCC,"System field: Supermicro SYS-422GA-NBRT-LCC (8x B200-SXM-180GB, TensorRT). Processor AMD TURIN 9965. Platform vSYS-422GA-NBRT-B200-SXM-180GBx8_TRT",CANNOT_TELL,"The entry names SYS-422GA-NBRT-LCC and lists the AMD processor that the post gives for the virtual system. The post names AS-4126GS-NBR-LCC, which is a separate Supermicro bare metal entry (5.1-0086).",false,Broadcom VMware Cloud Foundation Blog: The New Paradigm: MLPerf Inference 5.1 Confirms VCF is the Future of AI/ML Performance (post dated 15 December 2025),https://blogs.vmware.com/cloud-foundation/2025/12/15/mlperf-5-1-confirms-vcf-future-of-ai-ml-performance/,2026-10-28,MLPerf Inference v5.1
+GV2,Google,Google v5.0,Google Cloud was the only cloud provider to submit results on NVIDIA HGX B200 GPUs,true,"submitters with B200 (not GB200) accelerators in raw/inf_5.0.json, by division",only cloud provider on HGX B200,Google closed: 10 rows. Lambda open: 3 rows. NVIDIA closed: 18 rows. Supermicro closed: 26 rows,ABOVE,"The file shows more cloud providers on B200 than the one the post states. In the closed division only Google, NVIDIA and Supermicro submitted on eight B200. Lambda, a cloud provider, has open division entries on eight B200 (5.0-5657, 5.0-5658, 5.0-5659). CoreWeave submitted GB200.",false,Google Cloud Blog: AI Hypercomputer inference updates for Google Cloud TPU and GPU (post dated April 2025),https://cloud.google.com/blog/products/compute/ai-hypercomputer-inference-updates-for-google-cloud-tpu-and-gpu,2026-10-28,MLPerf Inference v5.0
+GV1,Google,Google v5.0,"we submitted 15 results, including our first submission with A3 Ultra (NVIDIA H200) and A4 (NVIDIA HGX B200) VMs",true,"Google closed datacenter rows in raw/inf_5.0.json, counted four ways",15 results,"3 entries (5.0-0035, 5.0-0036, 5.0-0078). 11 system and benchmark pairs. 22 with scenarios. 30 with accuracy targets. 30 rows",CANNOT_TELL,"The post does not define a result. No counting rule gives 15. The 5.0-0078 entry is named NVIDIA DGX B200 in the file, and the post calls it the A4 VM.",false,Google Cloud Blog: AI Hypercomputer inference updates for Google Cloud TPU and GPU (post dated April 2025),https://cloud.google.com/blog/products/compute/ai-hypercomputer-inference-updates-for-google-cloud-tpu-and-gpu,2026-10-28,MLPerf Inference v5.0
+GV3,Google,Google v5.0,"Trillium now delivers 3.5x throughput improvement for queries/second on Stable Diffusion XL (SDXL) compared to last performance round for its predecessor, TPU v5e. This further improves throughput by 12% since the MLPerf 4.1 submission.",true,TPU Trillium SDXL rows in raw/inf_5.0.json,3.5x over v5e and 12% over v4.1,"Offline 6.079, Server 5.480",CANNOT_TELL,The v5.0 Trillium values are in the file (4 chips). The v4.1 results for TPU v5e and Trillium are not in the files used here.,false,Google Cloud Blog: AI Hypercomputer inference updates for Google Cloud TPU and GPU (post dated April 2025),https://cloud.google.com/blog/products/compute/ai-hypercomputer-inference-updates-for-google-cloud-tpu-and-gpu,2026-10-28,MLPerf Inference v5.0
+GV4,Google,Google v5.0,"Google Cloud's A3 Ultra demonstrated highly competitive performance, achieving results comparable to NVIDIA's peak GPU submissions across LLMs, MoE, image, and recommendation models",true,"Google 5.0-0035 (A3 Ultra, eight H200) against NVIDIA 5.0-0060 (eight H200), Offline",comparable to NVIDIA peak,"dlrm-v2-99 0.989, dlrm-v2-99.9 0.994, gptj-99 0.993, gptj-99.9 0.993, llama2-70b-99 0.976, llama2-70b-99.9 0.976, llama3.1-405b 0.986, mixtral-8x7b 0.984, retinanet 1.014",CANNOT_TELL,The post gives no threshold for comparable. The ratios run from 0.989 to 1.014. Peak is read as NVIDIA on the same eight H200.,false,Google Cloud Blog: AI Hypercomputer inference updates for Google Cloud TPU and GPU (post dated April 2025),https://cloud.google.com/blog/products/compute/ai-hypercomputer-inference-updates-for-google-cloud-tpu-and-gpu,2026-10-28,MLPerf Inference v5.0
+GV5,Google,Google v5.0,JetStream achieves 1703 token/s on Llama 3.1 405B on Trillium,true,"Google internal measurement, not an MLPerf entry",1703 token/s on Llama 3.1 405B,no Trillium Llama 3.1 405B entry in raw/inf_5.0.json,CANNOT_TELL,The post presents Figures 1 to 4 as Google internal data.,false,Google Cloud Blog: AI Hypercomputer inference updates for Google Cloud TPU and GPU (post dated April 2025),https://cloud.google.com/blog/products/compute/ai-hypercomputer-inference-updates-for-google-cloud-tpu-and-gpu,2026-10-28,MLPerf Inference v5.0
+KV4,K-Veritas,K-Veritas,each report lists sessions started but never sealed.,true,"five local-mode scenarios: an abandoned session A, then a second session B sealed and verified",a report lists every session started and not sealed,"2 of 2 scenarios refused to start B while A was open (control_abandon_A, delete_A_folder). 3 of 3 ended with ""1 started, 1 sealed, 0 never sealed"" with A absent (clean_A, delete_session_ledger, delete_signer_key_and_ledger)",BELOW,"kveritas clean is documented to erase a session so that no later report lists it. Deleting the local session file or the signer key does the same. The report prints ""New signer: no earlier history, other attempts cannot be ruled out"" in those cases, as it does for any first report. The sibling list for a hosted seal was not tested.",false,"K-Veritas README, kveritas-go commit 521070c (26 October 2026)",https://github.com/27-GROUP/kveritas-go,2026-10-28,kveritas-go local-mode tests
+KV5,K-Veritas,K-Veritas,"Declared FLOPs vs time, energy, memory. PASS / REVIEW / FABRICATION-IMPOSSIBLE / N/A.",true,two canary scripts that print the same result line and take 0.13 seconds,the compute check returns one of four verdicts,"Canary 1 declared a 7B model for 3 epochs of 1,000,000 examples and read Compute [FABRICATION-IMPOSSIBLE]. Canary 2 declared a 1,000-parameter model and 100 examples and printed no compute line (certificate not applicable, from the source), with the claim val_acc = 0.9713 sealed",CONSISTENT,"A result printed by a script that declares a workload below the compute floor earns no flag, which the tool describes as outside its scope (""Not that the experiment is correct"").",false,"K-Veritas README, kveritas-go commit 521070c (26 October 2026)",https://github.com/27-GROUP/kveritas-go,2026-10-28,kveritas-go local-mode tests
+KV1,K-Veritas,K-Veritas,"these numbers came from this code, at this time, unchanged.",true,"five edits to a sealed local-mode report, each verified offline",a sealed result cannot be changed without detection,"5 of 5 edits read TAMPERED: a metric value 0.8301 -> 0.9301. The invocation count 3 -> 1 (hides the failed attempt). A kept run exit code 0 -> 7. The visible page text 0.8301 -> 0.9301. The self-reported signer history, started 1 -> 9",CANNOT_TELL,"The edits were a metric value, the invocation count, a kept run exit code, the visible page text and the self-reported signer history. The report reads SELF-ATTESTED, which the tool prints as ""signed with an author-supplied key, not K-Veritas"". The edits test the word unchanged. The code and time parts of the sentence were not tested, so the row reads CANNOT_TELL.",false,"K-Veritas README, kveritas-go commit 521070c (26 October 2026)",https://github.com/27-GROUP/kveritas-go,2026-10-28,kveritas-go local-mode tests
+KV2,K-Veritas,K-Veritas,"every run is recorded, failed ones too, so an attempt shows.",true,"a local-mode session with two kept runs and one failed run, sealed and verified offline",a failed attempt shows in the report,"The failed run printed ""It is not kept in the report, but the server still counts it in the run history."" The verify output lists Runs: 2 and no line for the failed attempt. The visible pages name no failed attempt. The embedded record holds 2 runs and an invocation count of 3",CANNOT_TELL,"Verify prints a total of invocations only when a hosted seal supplies one (cmd/kveritas/main.go line 2305). The hosted service was not used for this row, so the sentence is not tested in the mode it describes. In local mode the failed run does not show.",false,"K-Veritas README, kveritas-go commit 521070c (26 October 2026)",https://github.com/27-GROUP/kveritas-go,2026-10-28,kveritas-go local-mode tests
+KV3,K-Veritas,K-Veritas,"Every run anchored at the server when it ends, failed ones too.",true,"needs the hosted service, which these tests did not use","every run, failed ones too, is anchored at the server",not run: the tests ran in local mode with the network blocked,CANNOT_TELL,"About ten sessions with a single print each were registered at the hosted service from this Mac by mistake before the network was blocked. A hosted registration sends a session id, a hashed machine id and run digests, with no script text.",false,"K-Veritas README, kveritas-go commit 521070c (26 October 2026)",https://github.com/27-GROUP/kveritas-go,2026-10-28,kveritas-go local-mode tests
+KV6,K-Veritas,K-Veritas,"on every run: time, CPU, memory, GPU, energy.",true,peak GPU memory recorded for a run that prints one line. Source read for how the counter is sampled on macOS,GPU is measured on every run,"A script that prints one line recorded peak GPU memory of about 12,400 MB and 100% GPU use. The source reads the system-wide ioreg counters ""In use system memory"" and ""Device Utilization %"" (internal/hardware/vendors.go lines 158 to 176)",CANNOT_TELL,"The sentence does not say whether GPU figures are per process or system-wide. Other jobs were using the GPU on this Mac during the tests, so the recorded figure is not the script's own. The compute check uses the same peak figure in its memory review.",false,"K-Veritas README, kveritas-go commit 521070c (26 October 2026)",https://github.com/27-GROUP/kveritas-go,2026-10-28,kveritas-go local-mode tests
+NB1,Nebius,Nebius Storage v3.0,"fed 768 simulated NVIDIA B200 accelerators running the RetinaNet training workload, the highest accelerator count of any submission in this round.",true,"num_accelerators in every Closed division summary.json, all categories",the highest accelerator count of any submission,"Nebius RetinaNet 768. Next RetinaNet 736 (UBIX). Largest training entry of any workload 768. Checkpointing entries list up to 1024 (Azure, llama3-1t)",BELOW,"The checkpointing rule 4.6.2 counts simulated accelerators as the number of processes, so the 1,024 in the Azure and Everpure llama3-1t summaries is a count of simulated accelerators in the file. The post limits itself to RetinaNet two sentences later (next row).",false,Nebius blog: MLPerf Storage v3.0 results: Leading RetinaNet training on Nebius Object Storage (2026),https://nebius.com/blog/posts/mlperf-storage-v3-0-results,2026-10-28,MLPerf Storage v3.0
+NB11,Nebius,Nebius Storage v3.0,"Client nodes ran as publicly available CPU-only virtual machines (the AMD EPYC Genoa cpu-d3 preset) on Ubuntu 24.04 over Ethernet, with capped memory.",true,system description yaml for the RetinaNet entry,"cpu-d3, AMD EPYC Genoa, Ubuntu 24.04, Ethernet","chassis model_name cpu-d3, cpu_model AMD EPYC-Genoa Processor, operating_system Ubuntu 24.04, networking type ethernet (400 Gb/s)",CONSISTENT,Read in full from storage/v3.0/closed/Nebius/systems/nebius-s3-retinanet-32nx24-cm64.yaml. Memory is listed as 61 GiB per client.,false,Nebius blog: MLPerf Storage v3.0 results: Leading RetinaNet training on Nebius Object Storage (2026),https://nebius.com/blog/posts/mlperf-storage-v3-0-results,2026-10-28,MLPerf Storage v3.0
+NB2,Nebius,Nebius Storage v3.0,"Nebius ran RetinaNet at the highest accelerator count of all submissions, regardless of storage architecture.",true,"RetinaNet Closed entries, accelerators",highest RetinaNet accelerator count of all submissions,"Nebius 768 against 736 (UBIX, file storage) and 675 (XSKY)",CONSISTENT,Accelerator counts are read from num_accelerators in the summary files.,false,Nebius blog: MLPerf Storage v3.0 results: Leading RetinaNet training on Nebius Object Storage (2026),https://nebius.com/blog/posts/mlperf-storage-v3-0-results,2026-10-28,MLPerf Storage v3.0
+NB3,Nebius,Nebius Storage v3.0,that throughput is more than twice the next-best result.,true,"RetinaNet object storage entries, samples per second (mean of the last five runs)",more than twice the next-best object storage result,"Nebius 341881 samples/s, next 166384 (NVIDIA), ratio 2.055. Accelerators 768 against 370, ratio 2.076",CONSISTENT,"The post says ""that throughput"", read as samples per second. Bandwidth in GiB/s gives 2.055.",false,Nebius blog: MLPerf Storage v3.0 results: Leading RetinaNet training on Nebius Object Storage (2026),https://nebius.com/blog/posts/mlperf-storage-v3-0-results,2026-10-28,MLPerf Storage v3.0
+NB4,Nebius,Nebius Storage v3.0,Nebius posted the highest Unet3D accelerator count among object storage submissions this round.,true,Unet3D Closed entries with an object benchmark API,highest Unet3D accelerator count among object storage submissions,"Nebius 21, next 20 (NVIDIA)",CONSISTENT,Object storage is read from benchmark_API in each system description.,false,Nebius blog: MLPerf Storage v3.0 results: Leading RetinaNet training on Nebius Object Storage (2026),https://nebius.com/blog/posts/mlperf-storage-v3-0-results,2026-10-28,MLPerf Storage v3.0
+NB6,Nebius,Nebius Storage v3.0,"sustained an aggregate 341,881 samples per second while maintaining 88.4% accelerator utilization.",true,"RetinaNet entry, mean of runs 2 to 6 (run 1 is the warm-up, Rules 2.1.17)","341,881 samples/s at 88.4% utilization","341881 samples/s at 88.4%, 102.83 GiB/s",CONSISTENT,"The mean of all six runs is lower. The post uses the five after the warm-up, as the rules say.",false,Nebius blog: MLPerf Storage v3.0 results: Leading RetinaNet training on Nebius Object Storage (2026),https://nebius.com/blog/posts/mlperf-storage-v3-0-results,2026-10-28,MLPerf Storage v3.0
+NB7,Nebius,Nebius Storage v3.0,"21 simulated B200 accelerators across 21 client nodes, and Object Storage sustained 114.45 GiB/s of aggregate read throughput at 92.5% accelerator utilization.",true,"Unet3D entry, mean of runs 2 to 6","21 accelerators, 21 nodes, 114.45 GiB/s, 92.5%","21 accelerators, 21 hosts, 114.45 GiB/s, 92.5%",CONSISTENT,"Bandwidth is the file value in MiB/s divided by 1,024.",false,Nebius blog: MLPerf Storage v3.0 results: Leading RetinaNet training on Nebius Object Storage (2026),https://nebius.com/blog/posts/mlperf-storage-v3-0-results,2026-10-28,MLPerf Storage v3.0
+NB8,Nebius,Nebius Storage v3.0,Checkpoint write bandwidth scaled from 2.81 GiB/s on one client node to 19.46 GiB/s on eight.,true,Nebius llama3-8b and llama3-70b checkpointing summaries,"2.81 to 19.46 GiB/s write, one to eight client nodes","2.81 GiB/s on 1 host, 19.46 GiB/s on 8 hosts",CONSISTENT,The file field is named GB per second. The post and the rules call the unit GiB/s.,false,Nebius blog: MLPerf Storage v3.0 results: Leading RetinaNet training on Nebius Object Storage (2026),https://nebius.com/blog/posts/mlperf-storage-v3-0-results,2026-10-28,MLPerf Storage v3.0
+NB9,Nebius,Nebius Storage v3.0,aggregate read bandwidth increased from 7.14 GiB/s to 38.27 GiB/s.,true,"same two summaries, load_checkpoint_io_mean",7.14 to 38.27 GiB/s read,7.14 and 38.27 GiB/s,CONSISTENT,,false,Nebius blog: MLPerf Storage v3.0 results: Leading RetinaNet training on Nebius Object Storage (2026),https://nebius.com/blog/posts/mlperf-storage-v3-0-results,2026-10-28,MLPerf Storage v3.0
+NB10,Nebius,Nebius Storage v3.0,All four Closed division results ran on Object Storage with the Enhanced storage class in one of our production regions.,true,Nebius Closed entries and their system descriptions,"four Closed results, all on Object Storage with the Enhanced storage class","4 Closed entries (llama3-70b, llama3-8b, retinanet, unet3d), benchmark API object, URI scheme s3",CANNOT_TELL,The count and the object API match. A storage class and a production region are not fields in the system files.,false,Nebius blog: MLPerf Storage v3.0 results: Leading RetinaNet training on Nebius Object Storage (2026),https://nebius.com/blog/posts/mlperf-storage-v3-0-results,2026-10-28,MLPerf Storage v3.0
+NB12,Nebius,Nebius Storage v3.0,entry v3.0-0074.,true,"public entry IDs are on the MLCommons results page, not in the results repository",RetinaNet is entry v3.0-0074,the repository holds no entry IDs,CANNOT_TELL,The repository is organized by submitter and system name.,false,Nebius blog: MLPerf Storage v3.0 results: Leading RetinaNet training on Nebius Object Storage (2026),https://nebius.com/blog/posts/mlperf-storage-v3-0-results,2026-10-28,MLPerf Storage v3.0
+NB5,Nebius,Nebius Storage v3.0,only three benchmarked an S3 object storage backend.,true,"Closed submitters with an object benchmark API in a system description, and the storage URI scheme where the file names one",three of 19 Closed submitters used an S3 object storage backend,"19 Closed submitters. Object benchmark API: NVIDIA, Nebius, OpenLake. The file names an s3 scheme for: Nebius",CANNOT_TELL,The count of three matches. Whether the other two use an S3 interface is not stated in their system files.,false,Nebius blog: MLPerf Storage v3.0 results: Leading RetinaNet training on Nebius Object Storage (2026),https://nebius.com/blog/posts/mlperf-storage-v3-0-results,2026-10-28,MLPerf Storage v3.0
+AL1,Azure,Azure Lustre Storage v3.0,All 17 Azure results use Azure Managed Lustre as a shared remote file system through the POSIX protocol.,true,Azure Closed rows in the results repository and their system descriptions,"17 results, remote shared file system, POSIX","17 Azure rows (13 training, 4 checkpointing). storage_location remote, benchmark API file",CONSISTENT,"The system files list a file API, not the word POSIX.",false,Microsoft Learn: Azure Managed Lustre MLPerf Storage v3.0 results (2026),https://learn.microsoft.com/en-us/azure/azure-managed-lustre/mlperf-storage-v3-results,2026-10-28,MLPerf Storage v3.0
+AL2,Azure,Azure Lustre Storage v3.0,The Azure runs maintained more than 90% simulated accelerator utilization for five consecutive hours.,true,"accelerator utilization (mean of runs 2 to 6) and the start and end times of those five runs, all 13 training entries",more than 90% utilization for five consecutive hours,lowest utilization 91.6% of 13 entries. The five timed runs add up to 4.82 to 5.21 hours,CONSISTENT,"Each run lasts about an hour and the five scored runs are consecutive with a few minutes between them. The sum of run time is 4.82 hours in the shortest entry, 3.6 percent under five hours. The rule used on MODEL for a rounded figure is up to 5 percent.",false,Microsoft Learn: Azure Managed Lustre MLPerf Storage v3.0 results (2026),https://learn.microsoft.com/en-us/azure/azure-managed-lustre/mlperf-storage-v3-results,2026-10-28,MLPerf Storage v3.0
+AL3,Azure,Azure Lustre Storage v3.0,"Across the configurations with three simulated accelerators, measured read bandwidth was between 17.3 and 17.4 GiB/s.",true,"3 accelerator entries, mean of runs 2 to 6, MiB/s divided by 1,024",17.3 to 17.4 GiB/s,17.31 to 17.37 GiB/s over 5 entries,CONSISTENT,,false,Microsoft Learn: Azure Managed Lustre MLPerf Storage v3.0 results (2026),https://learn.microsoft.com/en-us/azure/azure-managed-lustre/mlperf-storage-v3-results,2026-10-28,MLPerf Storage v3.0
+AL4,Azure,Azure Lustre Storage v3.0,"At 16 simulated accelerators, measured read bandwidth was between 89.0 and 90.6 GiB/s across all five Azure Managed Lustre service tiers.",true,16 accelerator entries,89.0 to 90.6 GiB/s,89.02 to 90.64 GiB/s over 5 entries,CONSISTENT,,false,Microsoft Learn: Azure Managed Lustre MLPerf Storage v3.0 results (2026),https://learn.microsoft.com/en-us/azure/azure-managed-lustre/mlperf-storage-v3-results,2026-10-28,MLPerf Storage v3.0
+AL5,Azure,Azure Lustre Storage v3.0,"Across the five tested tiers, demonstrated density was between 95.6% and 97.3% of nominal.",true,"measured GiB/s over nominal GiB/s = capacity / increment x 2,000 / 1,073.741824, the page formula, for the five 16 accelerator entries",95.6% to 97.3%,"95.6% to 97.3% (95.6, 96.9, 97.0, 97.2, 97.3)",CONSISTENT,Density over nominal density is the same ratio as bandwidth over nominal bandwidth.,false,Microsoft Learn: Azure Managed Lustre MLPerf Storage v3.0 results (2026),https://learn.microsoft.com/en-us/azure/azure-managed-lustre/mlperf-storage-v3-results,2026-10-28,MLPerf Storage v3.0
+AL6,Azure,Azure Lustre Storage v3.0,The AMLFS 125 configurations increased from 3 simulated accelerators and 17.4 GiB/s (18.6 GB/s) to 70 simulated accelerators and 379.1 GiB/s (407.1 GB/s).,true,"AMLFS 125 entries at 3 and 70 accelerators, GB/s = GiB/s x 1.073741824",17.4 (18.6) at 3 and 379.1 (407.1) at 70,17.37 GiB/s (18.65 GB/s) at 3 and 379.14 (407.10) at 70,CONSISTENT,"The GB/s figure at 3 accelerators computes to 18.65, which the page prints as 18.6.",false,Microsoft Learn: Azure Managed Lustre MLPerf Storage v3.0 results (2026),https://learn.microsoft.com/en-us/azure/azure-managed-lustre/mlperf-storage-v3-results,2026-10-28,MLPerf Storage v3.0
+AL7,Azure,Azure Lustre Storage v3.0,"In addition to the AMLFS 125 result, the AMLFS 20 configuration measured 378.0 GiB/s.",true,AMLFS 20 entry at 70 accelerators,378.0 GiB/s,378.02 GiB/s,CONSISTENT,,false,Microsoft Learn: Azure Managed Lustre MLPerf Storage v3.0 results (2026),https://learn.microsoft.com/en-us/azure/azure-managed-lustre/mlperf-storage-v3-results,2026-10-28,MLPerf Storage v3.0
+AL8,Azure,Azure Lustre Storage v3.0,"The Complete training results table on the page lists 13 Azure entries by tier, capacity, simulated accelerators and read bandwidth.",false,"the 13 rows of the page table matched to entries by tier, capacity and accelerator count in the system name",read bandwidth in GiB/s for 13 entries,13 of 13 rows match to one decimal,CONSISTENT,"Boundary values such as 90.35 and 90.25 round either way at one decimal, so a difference of up to 0.05 is accepted.",false,Microsoft Learn: Azure Managed Lustre MLPerf Storage v3.0 results (2026),https://learn.microsoft.com/en-us/azure/azure-managed-lustre/mlperf-storage-v3-results,2026-10-28,MLPerf Storage v3.0
+AL9,Azure,Azure Lustre Storage v3.0,"The largest result, representing the 1250B model, measured 642.2 GiB/s (689.6 GB/s) for checkpoint writes and 489.6 GiB/s (525.7 GB/s) for recovery reads.",true,the four checkpointing entries: write and read GiB/s and durations compared with the page table,"four rows: 8B, 70B, 405B and 1250B",4 of 4 rows match. The 642.2 and 489.6 values are the llama3-1t entry,CONSISTENT,"The results repository names the largest checkpoint workload llama3-1t, with a checkpoint of 15,426 GB. The page labels it 1250B. The file holds no parameter count, so the label is not checked.",false,Microsoft Learn: Azure Managed Lustre MLPerf Storage v3.0 results (2026),https://learn.microsoft.com/en-us/azure/azure-managed-lustre/mlperf-storage-v3-results,2026-10-28,MLPerf Storage v3.0
+AL10,Azure,Azure Lustre Storage v3.0,The Public ID column on the page lists entries from v3.0-0031 to v3.0-0047.,false,"public entry IDs are on the MLCommons results page, not in the results repository",entries v3.0-0031 to v3.0-0047,the repository holds no entry IDs,CANNOT_TELL,"The 17 entries match the page table by tier, capacity, accelerators and values. The ID column is not checked.",false,Microsoft Learn: Azure Managed Lustre MLPerf Storage v3.0 results (2026),https://learn.microsoft.com/en-us/azure/azure-managed-lustre/mlperf-storage-v3-results,2026-10-28,MLPerf Storage v3.0
+AS1,NVIDIA,NVIDIA AIStore Storage v3.0,NVIDIA published 20 Closed Division results across training and checkpointing workloads.,true,Closed division entries submitted by NVIDIA in the results repository,20 Closed Division results,20 Closed entries. Categories checkpointing and training,CONSISTENT,Counted from the results repository.,false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS10,NVIDIA,NVIDIA AIStore Storage v3.0,"The benchmark delivered 136.54 GiB/s of recovery-read throughput and 133.27 GiB/s of write throughput, equivalent to 97.7% and 95.4% of that nominal figure, respectively.",true,"Llama 3 1T summary at 12 nodes, read and write, divided by 12 times 11.64 GiB/s","136.54 read, 133.27 write, 97.7% and 95.4%","136.54 read, 133.27 write, 97.7% and 95.4%",CONSISTENT,Percentages use the exact nominal 139.70 GiB/s.,false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS11,NVIDIA,NVIDIA AIStore Storage v3.0,"RetinaNet has roughly 450 times as many objects per GiB as UNet3D, whose average sample is 140 MiB.",true,the post's own sizes: 140 MiB per UNet3D sample and about 315 KiB per RetinaNet object,roughly 450 times,455 times,CONSISTENT,"The object sizes are from the post, not the results files. The arithmetic is checked.",false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS12,NVIDIA,NVIDIA AIStore Storage v3.0,"The 1T-parameter model checkpoint runs used 8, 8, and 16 clients, respectively",true,"client nodes (hosts) in the Llama 3 1T entries at 3, 6 and 12 AIStore nodes","8, 8 and 16 clients","8, 8 and 16 client nodes",CONSISTENT,"The rest of the sentence, two data-parallel instances at every scale, is not in the fields read here.",false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS13,NVIDIA,NVIDIA AIStore Storage v3.0,The training runs scaled from 5 to 10 to 20 clients and simulated B200 accelerators,true,"UNet3D entries at 3, 6 and 12 AIStore nodes, client nodes and simulated accelerators","5, 10 and 20 clients and accelerators","hosts 5, 10, 20. Accelerators 5, 10, 20",CONSISTENT,Read from the three UNet3D entries.,false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS15,NVIDIA,NVIDIA AIStore Storage v3.0,"the submission is classified as Available, meaning all system components are publicly available, and as Shared Remote Object storage: multiple clients access the storage over a network through an object API.",true,"availability, storage location and benchmark API on all NVIDIA Closed entries",Available. Shared Remote Object storage,"availability available, storage location remote, benchmark API object",CONSISTENT,Read from the entry classifications.,false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS16,NVIDIA,NVIDIA AIStore Storage v3.0,It was submitted to the Closed Division,true,division of NVIDIA entries,Closed Division,divisions closed,CONSISTENT,All 20 entries sit in the Closed folder.,false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS2,NVIDIA,NVIDIA AIStore Storage v3.0,Increasing the AIStore cluster from 3 to 12 nodes (4x) delivered 3.97x the UNet3D training I/O and 3.99x the Llama 3 1T checkpoint recovery throughput,true,UNet3D I/O at 12 nodes over 3 nodes. Llama 3 1T load_checkpoint_io_mean at 12 nodes over 3 nodes,3.97x and 3.99x,UNet3D 115.58 over 29.15 = 3.966x. 1T read 136.54 over 34.20 = 3.993x. The 1T runs used 8 client nodes at 3 AIStore nodes and 16 at 12.,CONSISTENT,"Both ratios round to the stated figures. The client count was doubled for the 12 node checkpoint run, which the post states in its table.",false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS3,NVIDIA,NVIDIA AIStore Storage v3.0,"UNet3D throughput remained nearly constant per storage node: 9.72 GiB/s at three nodes, 9.70 GiB/s at six, and 9.63 GiB/s at twelve.",true,UNet3D I/O divided by AIStore node count,"9.72, 9.70 and 9.63 GiB/s per node","9.72, 9.70 and 9.63 GiB/s per node",CONSISTENT,"Node counts taken from the submission names (3n, 6n, 12n).",false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS4,NVIDIA,NVIDIA AIStore Storage v3.0,"From three to six OCI nodes, RetinaNet I/O increased from 27.09 to 50.04 GiB/s, a 1.85x gain.",true,RetinaNet I/O at 3 and 6 OCI nodes,"27.09 to 50.04 GiB/s, 1.85x","27.09 to 50.04 GiB/s, 1.848x",CONSISTENT,"Mean of the last five runs, as in every row.",false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS5,NVIDIA,NVIDIA AIStore Storage v3.0,"Mean AU was 94.06% at three nodes and 89.25% at six nodes, both above the workload's 85% requirement.",true,"RetinaNet accelerator utilization at 3 and 6 OCI nodes, against the 85% requirement stated in the post","94.06% and 89.25%, both above 85%",94.06% and 89.25%,CONSISTENT,The 85% requirement is as the post states it.,false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS6,NVIDIA,NVIDIA AIStore Storage v3.0,I/O per simulated B200 stayed in a narrow range of 5.77 to 5.83 GiB/s.,true,UNet3D I/O divided by simulated accelerators on the three cloud deployments,5.77 to 5.83 GiB/s,"AWS 5.80, Google Cloud 5.77, OCI 5.83 GiB/s per accelerator",CONSISTENT,Accelerators are the simulated accelerator count in each entry.,false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS7,NVIDIA,NVIDIA AIStore Storage v3.0,Three separate cloud deployments sustained more than 97% mean AU through the S3 object API.,true,"UNet3D accelerator utilization on the AWS, Google Cloud and OCI entries, and their benchmark API","more than 97% mean AU, object API","98.37%, 97.88%, 98.86%. Benchmark API object",CONSISTENT,The files record an object benchmark API. Whether it is S3 is as the post states.,false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS8,NVIDIA,NVIDIA AIStore Storage v3.0,"Each storage node had a nominal 100 Gb/s network interface, equivalent to 11.64 GiB/s before protocol overhead.",true,100 Gb/s converted to GiB/s,11.64 GiB/s,11.642 GiB/s,CONSISTENT,Arithmetic on the stated interface speed. The interface itself is not in the results files.,false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS9,NVIDIA,NVIDIA AIStore Storage v3.0,"Across twelve nodes, that provided 139.7 GiB/s of nominal aggregate NIC bandwidth.",true,12 nodes at 100 Gb/s each,139.7 GiB/s,139.70 GiB/s,CONSISTENT,Arithmetic on the stated interface speed.,false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS14,NVIDIA,NVIDIA AIStore Storage v3.0,Our submission used NVIDIA AIStore v4.8 on Kubernetes.,true,NVIDIA system description files: AIStore version and Kubernetes,AIStore v4.8 on Kubernetes,17 system files. AIStore v4.8 appears in the description of 17. Kubernetes appears in 16.,CANNOT_TELL,"Counted from the yaml files. Where a file names Google Kubernetes Engine it counts as Kubernetes. One file, aistore-3n-oci-bm-denseio-e5-128-1client.yaml, does not name an orchestrator. The other 16 do, so the record is silent on that one entry and contradicts none.",false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+AS17,NVIDIA,NVIDIA AIStore Storage v3.0,Cloud object storage was not part of the measured data path.,true,"the results repository describes the system under test, not a negative about other storage",cloud object storage not in the measured data path,the files list AIStore nodes as the storage and name no other,CANNOT_TELL,A system description lists what was used. It does not state what was absent.,false,NVIDIA AIStore blog: NVIDIA AIStore at MLPerf Storage v3.0: Linear Scale-Out and Multi-Cloud Portability (2026),https://aistore.nvidia.com/blog/2026/09/01/mlperf-storage-v3,2026-10-28,MLPerf Storage v3.0
+T1,NVIDIA,NVIDIA Training v6.0,"Table 1 in the post lists DeepSeek-V3 671B (MoE) on GB300 NVL72 with 8,192 GPUs at 2.02 mins.",false,"run_stop minus run_start of every result log in the entry, olympic average (drop highest and lowest), times the entry's scaling factor. Fastest GB300 entry with 8,192 GPUs",2.02 minutes,"2.021 minutes from runs 1.965, 1.945, 1.891 (CoreWeave)",CONSISTENT,"The entry is CoreWeave's CoreWeave_GB300_2048x4. A scaling factor of 1.0389 set by the review committee is applied, as in the official summarizer.",false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T10,NVIDIA,NVIDIA Training v6.0,By securing a clean sweep and winning every single benchmark in this round,true,"the fastest time to train per benchmark, computed as in T1 to T7, across every result set",fastest time on every benchmark,NVIDIA accelerators hold the fastest time on 7 of 7 benchmarks,CONSISTENT,"Time to train depends on the number of GPUs. The fastest entries are at 64 to 8,192 GPUs.",false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T11,NVIDIA,NVIDIA Training v6.0,also delivered the highest performance when normalized on a per-accelerator basis on every benchmark,true,"best GPU-minutes (time to train times GPU count, lower is better) per benchmark, NVIDIA accelerators against others",highest per-accelerator performance on every benchmark,"flux1 1,799.1 against 5,911.0. llama2_70b_lora 44.9 against 66.2. llama31_8b 494.9 against 694.8. No other vendor entries on 4 benchmarks",CONSISTENT,GPU-minutes is our own normalization. The post does not define the per-accelerator measure.,false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T12,NVIDIA,NVIDIA Training v6.0,"NVIDIA cloud service provider partners scaled up to 8,192 Blackwell GPUs working in unison",true,"result sets with 8,192 GPUs","up to 8,192 Blackwell GPUs","Azure llama31_405b (GB200), CoreWeave deepseekv3_671b (GB300), NVIDIA deepseekv3_671b (GB200). Largest GPU count in any result set: 8,192",CONSISTENT,"Three result sets reach 8,192 GPUs. The largest GPU count in the repository is checked in the same step.",false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T14,NVIDIA,NVIDIA Training v6.0,"improved 1.3x going from 1,298 TFLOPS/GPU to 1,648 TFLOPS/GPU",true,"1,648 over 1,298, both numbers from the post",1.3x,1.270x,CONSISTENT,Arithmetic on the post's own numbers. The TFLOPS figures come from NVIDIA container tests and are not in the MLCommons logs.,false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T2,NVIDIA,NVIDIA Training v6.0,Table 1 in the post lists GPT-OSS 20B (MoE) on GB300 NVL72 with 512 GPUs at 7.43 mins.,false,"run_stop minus run_start of every result log in the entry, olympic average (drop highest and lowest), times the entry's scaling factor. Fastest GB300 entry with 512 GPUs",7.43 minutes,"7.431 minutes from runs 7.160, 7.723, 6.947, 7.188, 7.487, 7.218, 7.120, 7.225, 7.572, 7.569 (NVIDIA)",CONSISTENT,"The entry is NVIDIA's theia_ngpu512_ngc26.04_nemo. A scaling factor of 1.0155 set by the review committee is applied, as in the official summarizer. The log field submission_platform reads 128xNVIDIA GB200 while the system description lists NVIDIA Blackwell Ultra GPU (GB300).",false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T3,NVIDIA,NVIDIA Training v6.0,"Table 1 in the post lists Llama 3.1 405B on GB200 NVL72 with 8,192 GPUs at 7.07 mins.",false,"run_stop minus run_start of every result log in the entry, olympic average (drop highest and lowest), times the entry's scaling factor. Fastest GB200 entry with 8,192 GPUs",7.07 minutes,"7.069 minutes from runs 7.212, 6.642, 7.069 (Azure)",CONSISTENT,The entry is Azure's 2048xAzure_GB200.,false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T4,NVIDIA,NVIDIA Training v6.0,"Table 1 in the post lists Llama 3.1 8B on GB200 NVL72 with 1,024 GPUs at 4.46 mins.",false,"run_stop minus run_start of every result log in the entry, olympic average (drop highest and lowest), times the entry's scaling factor. Fastest GB200 entry with 1,024 GPUs",4.46 minutes,"4.459 minutes from runs 4.328, 4.074, 4.550, 4.546, 4.811, 4.374, 4.370, 4.371, 4.373, 4.626 (NVIDIA)",CONSISTENT,"The entry is NVIDIA's tyche-hsg_ngpu1024_ngc26.04_nemo. A scaling factor of 1.0038 set by the review committee is applied, as in the official summarizer.",false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T5,NVIDIA,NVIDIA Training v6.0,Table 1 in the post lists Llama 2 70B LoRA on GB300 NVL72 with 512 GPUs at 0.4 mins.,false,"run_stop minus run_start of every result log in the entry, olympic average (drop highest and lowest), times the entry's scaling factor. Fastest GB300 entry with 512 GPUs",0.4 minutes,"0.400 minutes from runs 0.521, 0.365, 0.367, 0.388, 0.388, 0.364, 0.408, 0.409, 0.411, 0.410 (NVIDIA)",CONSISTENT,"The entry is NVIDIA's theia_ngpu512_ngc26.04_nemo. A scaling factor of 1.0176 set by the review committee is applied, as in the official summarizer. The log field submission_platform reads 128xNVIDIA GB200 while the system description lists NVIDIA Blackwell Ultra GPU (GB300).",false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T6,NVIDIA,NVIDIA Training v6.0,Table 1 in the post lists FLUX.1 on GB300 NVL72 with 512 GPUs at 17.1 mins.,false,"run_stop minus run_start of every result log in the entry, olympic average (drop highest and lowest), times the entry's scaling factor. Fastest GB300 entry with 512 GPUs",17.1 minutes,"17.119 minutes from runs 16.791, 16.053, 16.815, 19.382, 16.078, 16.802, 17.206, 17.307, 17.625, 18.324 (NVIDIA)",CONSISTENT,The entry is NVIDIA's theia_ngpu512_ngc25.09_nemo. The log field submission_platform reads 128xNVIDIA GB200 while the system description lists NVIDIA Blackwell Ultra GPU (GB300).,false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T7,NVIDIA,NVIDIA Training v6.0,Table 1 in the post lists DLRM-dcnv2 on GB300 NVL72 with 64 GPUs at 0.67 mins.,false,"run_stop minus run_start of every result log in the entry, olympic average (drop highest and lowest), times the entry's scaling factor. Fastest GB300 entry with 64 GPUs",0.67 minutes,"0.672 minutes from runs 0.623, 0.665, 0.699, 0.663, 0.663, 0.625, 0.734, 0.737, 0.663, 0.629 (GigaComputing)",CONSISTENT,"The entry is GigaComputing's DLB2-CB3_hugectr. A scaling factor of 1.0068 set by the review committee is applied, as in the official summarizer.",false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T8,NVIDIA,NVIDIA Training v6.0,The NVIDIA platform was the only one to submit results on both new workloads,true,the accelerator model in every result set for deepseekv3_671b and gpt_oss_20b in the v6.0 repository,only NVIDIA on both new workloads,deepseekv3_671b: NVIDIA. gpt_oss_20b: NVIDIA,CONSISTENT,Both new workloads have result sets only on NVIDIA accelerators. Other vendors submitted on three of the seven benchmarks.,false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T9,NVIDIA,NVIDIA Training v6.0,It was also the only platform to submit on every test,true,benchmarks with result sets by accelerator vendor,only NVIDIA on every test,"NVIDIA on 7 of 7 benchmarks, another vendor on 3: flux1, llama2_70b_lora, llama31_8b",CONSISTENT,Counts are benchmarks present in the v6.0 result repository.,false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T13,NVIDIA,NVIDIA Training v6.0,"Figure 2 in the post shows a 1.3x performance gain for GPT-OSS 20B, a 1.5x gain for Llama 3.1 405B, and a 1.6x gain for DeepSeek-V3 from GB300 over GB200.",false,GB200 over GB300 time to train at the same GPU count in the v6.0 logs,"1.3x, 1.5x and 1.6x","same-count ratios in v6.0: DeepSeek-V3 1.42 to 1.65 at five scales (1.65 at 8,192 GPUs), GPT-OSS 20B 1.06 to 1.25 at four scales, Llama 3.1 405B no GB200 and GB300 pair at the same count",CANNOT_TELL,The footnote cites entries from v5.1 and v6.0 and the figure does not state the scale. The v6.0 logs alone cannot place the Llama 3.1 405B bar.,false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+T15,NVIDIA,NVIDIA Training v6.0,more than 8% end-to-end benefit on Deepseek-v3 and a 93% end-to-end speedup on GPT-OSS,true,NVIDIA's own before and after tests,more than 8% and 93%,not in the MLCommons logs,CANNOT_TELL,These are NVIDIA's internal comparisons of software versions.,false,NVIDIA Technical Blog: NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance,https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/,2026-10-28,MLPerf Training v6.0
+A1,AMD,AMD v5.1,MI355X shows 2.7 x improvement over MI325X due to superior architecture and MXFP4 capabilities,true,"5.1-0098 (open, 8xMI355X, MXFP4) / 5.1-0002 (closed, 8xMI325X, FP8), Llama 2 70B Offline, tokens/s",2.7x,2.695x (93045.8 vs 34520.4). Against the v5.0 entry 5.0-0001: 2.742x,CONSISTENT,"Open division against closed division and MXFP4 against FP8, as the blog's own figure caption says.",false,ROCm Blogs: Technical Dive into AMD's MLPerf Inference v5.1 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v5.1/README.html,2026-10-28,MLPerf Inference v5.1
+A10,AMD,AMD v5.1,we improved performance by approximately 8% in the offline scenario compared to our MLPerf Inference v5.0 submission,true,"AMD 8xMI325x SD-XL Offline, 5.1-0004 / 5.0-0002, samples/s",+8%,+8.7% (18.5856 vs 17.0963),CONSISTENT,within one point of the stated figure,false,ROCm Blogs: Technical Dive into AMD's MLPerf Inference v5.1 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v5.1/README.html,2026-10-28,MLPerf Inference v5.1
+A11,AMD,AMD v5.1,Note the performance leadership of our Mixtral-8x7B Offline score,true,"closed Mixtral-8x7B Offline per accelerator, MI325X (n=6) against H200 (n=2)","leadership over the H200 average (Figure 4, bars not in the text)","MI325X median 8563 tokens/s per GPU, H200 mean 7718 (1.109x). Only 2 H200 submissions",CONSISTENT,The figure values are not in the page text. Direction and size from the file.,false,ROCm Blogs: Technical Dive into AMD's MLPerf Inference v5.1 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v5.1/README.html,2026-10-28,MLPerf Inference v5.1
+A3,AMD,AMD v5.1,"Compared to the 4 node submission on MI325X in the previous round, submission ID 5.0-0054, our current MI355X 4 node score, submission ID 5.1-0102, is 3.4x faster",true,"5.1-0102 / 5.0-0054, Llama 2 70B Offline",3.4x,"3.400x (350820 vs 103182 tokens/s). The file lists entry 5.0-0054 as AMD Instinct MI300X 192GB HBM3, not MI325X",CONSISTENT,"Ratio holds. The entry named in the blog is a 32-GPU MI300X system in the file, and no multi-node MI325X entry exists in v5.0. 5.1-0102 is open division, 5.0-0054 is closed.",true,ROCm Blogs: Technical Dive into AMD's MLPerf Inference v5.1 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v5.1/README.html,2026-10-28,MLPerf Inference v5.1
+A4,AMD,AMD v5.1,"we have improved the performance by about 2% and 4% in offline and server scenarios, respectively, over MLPerf Inference v5.0 submission",true,"AMD entries 5.1-0002 / 5.0-0001, 8xMI325X, Llama 2 70B, closed","+2% offline, +4% server","+1.7% offline, +4.2% server",CONSISTENT,rounds to the stated figures,false,ROCm Blogs: Technical Dive into AMD's MLPerf Inference v5.1 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v5.1/README.html,2026-10-28,MLPerf Inference v5.1
+A5,AMD,AMD v5.1,94% efficiency compared to single-node submissions from these platforms,true,"5.1-0065 (4 MI300X nodes + 2 MI325X nodes, closed) / (4 x MI300X single node + 2 x MI325X single node), Llama 2 70B Offline",94%,"94.6% using median single-node rows (93.7% to 96.8% over all published single-node rows, 4 MI300X and 9 MI325X rows)",CONSISTENT,The blog does not say which single-node rows it used. The range over every published single-node row includes the stated figure.,false,ROCm Blogs: Technical Dive into AMD's MLPerf Inference v5.1 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v5.1/README.html,2026-10-28,MLPerf Inference v5.1
+A7,AMD,AMD v5.1,We obtain a performance of 1942 tokens/s ... we see a performance of 2019 tokens/s,true,"5.1-0101 and 5.1-0100, Llama 3.1 405B Offline, open, 8xMI355X",1942 and 2019 tokens/s,1942.15 and 2019.04 tokens/s,CONSISTENT,matches to rounding,false,ROCm Blogs: Technical Dive into AMD's MLPerf Inference v5.1 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v5.1/README.html,2026-10-28,MLPerf Inference v5.1
+A12,AMD,AMD v5.1,Figures 4 and 5 in the post compare MI325X results against an average of H200 results.,false,bar values are images on the page,figure bars,not readable by script,CANNOT_TELL,Values exist only in the figure images.,false,ROCm Blogs: Technical Dive into AMD's MLPerf Inference v5.1 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v5.1/README.html,2026-10-28,MLPerf Inference v5.1
+A2,AMD,AMD v5.1,"Compared to our internal baselines, the 4 node achieves 3.9x better performance and the 8 node submission shows 7.9x improvement over our single node baseline",true,"5.1-0102 and 5.1-0104 (AMD_MangoBoost, open) / 5.1-0098 (the only published single-node MI355X Llama 2 70B Offline row)",3.9x and 7.9x,3.77x and 6.97x against the published single-node row (93045.8 tokens/s). The stated ratios imply baselines of 89954 and 82057 tokens/s,CANNOT_TELL,The baseline is called internal and is not a published row. Against the one published single-node row the ratios are lower.,false,ROCm Blogs: Technical Dive into AMD's MLPerf Inference v5.1 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v5.1/README.html,2026-10-28,MLPerf Inference v5.1
+A6,AMD,AMD v5.1,partners delivering performance that is nearly identical to AMD reference results across both Instinct MI325X and MI300X platforms,true,"single-node closed Llama 2 70B rows, partner / AMD reference (5.1-0002 MI325X, 5.1-0001 MI300X)",nearly identical (no figure),"MI325X Offline 94.9% to 99.9% (n=7), Server 94.7% to 100.3%. MI300X Offline 98.3% to 100.2% (n=3), Server 99.8% to 101.1%",CANNOT_TELL,"No figure is stated, so there is nothing to hold the spread against. The spread is shown.",false,ROCm Blogs: Technical Dive into AMD's MLPerf Inference v5.1 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v5.1/README.html,2026-10-28,MLPerf Inference v5.1
+A8,AMD,AMD v5.1,A 21% depth-pruned model delivered up to an 82% increase in inference throughput ... 33% pruned model ... 90% performance uplift,true,no unpruned MI355X Llama 3.1 405B row exists in v5.1,+82% and +90%,no baseline row in the file,CANNOT_TELL,The uplift is against an unpruned model that was not submitted.,false,ROCm Blogs: Technical Dive into AMD's MLPerf Inference v5.1 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v5.1/README.html,2026-10-28,MLPerf Inference v5.1
+A9,AMD,AMD v5.1,The above optimizations resulted in a 23% end-to-end throughput improvement,true,Mixtral-8x7B kernel optimizations,+23%,no before and after rows in the file,CANNOT_TELL,The baseline is internal.,false,ROCm Blogs: Technical Dive into AMD's MLPerf Inference v5.1 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v5.1/README.html,2026-10-28,MLPerf Inference v5.1
+V4,AMD,AMD v6.0,This was accomplished with efficient scale-up and scale-out end-to-end ROCm solution stack optimization that achieved a 97-98% scaling efficiency.,true,the five scaling figures in V2 and V3,97-98%,"92.4% to 98.1%, with only Llama 2 70B interactive at 98.1%",BELOW,"The same post gives 93%, 98%, 92% and 93% for these runs.",false,ROCm Blogs: AMD Instinct GPUs MLPerf Inference v6.0 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v6.0/README.html,2026-10-28,MLPerf Inference v6.0
+V2,AMD,AMD v6.0,Offline and Server achieving 93% scaling efficiency while Interactive is even better at 98%.,true,"per GPU result of 6.0-0001 (87 GPUs) over per GPU result of the single node entry 6.0-0002 (8 GPUs), Llama 2 70B","93% offline and server, 98% interactive","92.6% offline, 93.2% server, 98.1% interactive",CONSISTENT,Uses 87 GPUs as the entry name states. The count in the file column Total Accelerators is 88.,false,ROCm Blogs: AMD Instinct GPUs MLPerf Inference v6.0 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v6.0/README.html,2026-10-28,MLPerf Inference v6.0
+V3,AMD,AMD v6.0,The results (see figure below) show an efficient distributed inference with scaling of 92% and 93% over 12 nodes.,true,"per GPU result of 6.0-0003 (94 GPUs) over per GPU result of 6.0-0002 (8 GPUs), gpt-oss-120b",92% and 93%,"92.4% offline, 93.3% server",CONSISTENT,Uses 94 GPUs as the entry name states. The count in the file column Total Accelerators is 96.,false,ROCm Blogs: AMD Instinct GPUs MLPerf Inference v6.0 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v6.0/README.html,2026-10-28,MLPerf Inference v6.0
+V5,AMD,AMD v6.0,"Against Nvidia’s B300, AMD’s GPU delivers 92% in offline mode, 93% in server mode, and exceeds with 104% in interactive mode.",true,"AMD 8 x MI355X (6.0-0002) over 8 x B300, per GPU, Llama 2 70B: NVIDIA entry 6.0-0073 offline and server, the best of the 2 B300 interactive rows (Dell) interactive","92% offline, 93% server, 104% interactive","91.6% offline, 93.4% server, 104.1% interactive",CONSISTENT,"Against the best B300 offline result from any submitter the ratio is 89.4%, and server 87.6%.",false,ROCm Blogs: AMD Instinct GPUs MLPerf Inference v6.0 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v6.0/README.html,2026-10-28,MLPerf Inference v6.0
+V6,AMD,AMD v6.0,achieving up to 115% in server mode and 111% in offline mode,true,"AMD 8 x MI355X (6.0-0002) over each of the 2 closed 8 x B200 gpt-oss-120b rows, per GPU","up to 115% server, 111% offline",against the lower B200 row 114.7% server (RedHat) and 110.6% offline (Nebius). Against the higher B200 row 93.9% server and 102.1% offline,CONSISTENT,The post says it shows the two best B200 results from NVIDIA partners because NVIDIA submitted none. Against the B300 rows the ratios are below 100%.,false,ROCm Blogs: AMD Instinct GPUs MLPerf Inference v6.0 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v6.0/README.html,2026-10-28,MLPerf Inference v6.0
+V7,AMD,AMD v6.0,achieved a throughput of 793 tokens/second,true,"entry 6.0-0102, Llama 3.1 405B Interactive, open division, 8 x MI355X",793 tokens/s,793.994 tokens/s,CONSISTENT,"Open division, with about 20% of the 126 layers dropped as the post says.",false,ROCm Blogs: AMD Instinct GPUs MLPerf Inference v6.0 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v6.0/README.html,2026-10-28,MLPerf Inference v6.0
+V9,AMD,AMD v6.0,"AMD had 9 partners: Oracle, Cisco, Dell, Giga Computing, MangoBoost, MiTAC, HPE, Supermicro, and Red Hat.",true,"submitters of v6.0 entries with an AMD Instinct accelerator, other than AMD",9 partners,"10 submitter names: Cisco, Dell, Dell_MangoBoost, GigaComputing, HPE, MangoBoost, MiTAC, ORACLE, Supermicro, Supermicro_RedHat",CONSISTENT,"Dell and MangoBoost also appear as one joint name, and Red Hat appears in the joint entry with Supermicro. That makes nine organizations.",false,ROCm Blogs: AMD Instinct GPUs MLPerf Inference v6.0 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v6.0/README.html,2026-10-28,MLPerf Inference v6.0
+V1,AMD,AMD v6.0,It was the first ever MLPerf submission with over a million tokens/sec for Llama 2 70B (11-nodes) and gpt-oss-120b (12-nodes) in the Closed division.,true,"entries 6.0-0001 (87 GPUs, 11 nodes) and 6.0-0003 (94 GPUs, 12 nodes), closed, tokens/s. Every other closed row at or over one million for those models in v5.0, v5.1 and v6.0","over a million tokens/s, first ever","Llama 2 70B 1,042,110 offline and 1,016,380 server. gpt-oss-120b 1,031,070 offline and 900,054 server. In the same round, NVIDIA 6.0-0078, Nebius 6.0-0081 also list closed results over one million for these models. No v5.0 or v5.1 row for them reaches one million (0 rows)",CANNOT_TELL,"The figures hold for the offline and server results except gpt-oss-120b server, which is under one million. The file has no submission order, so it cannot say which entry was first within the round.",false,ROCm Blogs: AMD Instinct GPUs MLPerf Inference v6.0 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v6.0/README.html,2026-10-28,MLPerf Inference v6.0
+V10,AMD,AMD v6.0,"The joint submission by Dell and MangoBoost is notable for being the first ever MLPerf Inference submission featuring three different GPUs: MI300X, MI325X, and MI355X.",true,"entry 6.0-0031, Llama 2 70B","three GPU types in one entry, first ever","entry 6.0-0031 lists AMD Instinct MI300X 192GB HBM3 (x8) and AMD Instinct MI325X 256GB HBM3e (x8) and AMD Instinct MI355X 288GB HBM3e (x8), 3 nodes",CANNOT_TELL,"The entry exists as described. Whether it is the first is outside the v5.0, v5.1 and v6.0 files.",false,ROCm Blogs: AMD Instinct GPUs MLPerf Inference v6.0 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v6.0/README.html,2026-10-28,MLPerf Inference v6.0
+V8,AMD,AMD v6.0,Offline scores improved by 4.4x and Server scores improved by 4.8x.,true,"6.0-0002 over the AMD 8 x MI325X entries of v5.0 (5.0-0001) and v5.1 (5.1-0002), Llama 2 70B","4.4x offline, 4.8x server","3.05x offline and 3.26x server over v5.0. 3.00x and 3.13x over v5.1. The stated ratios imply earlier scores of 23,518 and 20,892 tokens/s",CANNOT_TELL,"The comparison is shown in a figure against earlier rounds, and the earlier AMD result it uses is not in the v5.0 or v5.1 files.",false,ROCm Blogs: AMD Instinct GPUs MLPerf Inference v6.0 Submission,https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference-v6.0/README.html,2026-10-28,MLPerf Inference v6.0
+N1,Intel,Intel v5.1,performance per dollar advantages of up to 1.25x and up to 4x compared to NVIDIA RTX Pro 6000 and L40S respectively,true,"Llama 3.1 8B, closed: Intel entry 5.1-0056 against the only RTX PRO 6000 entry (5.1-0051) and the only L40S entry (5.1-0082) in the file",up to 1.25x and up to 4x performance per dollar,"Throughput only, no prices in the file. Intel 6,472.37 offline and 5,348.45 server tokens/s with 4 cards. RTX PRO 6000 entry 46,841.8 and 46,060.0 with 8 cards. L40S entry 1,642.22 and 1,207.14 with 1 card",CANNOT_TELL,"The post says the cost figures are Intel internal estimates from public and partner sourcing for one named configuration as of September 2025. The prices are not in the MLCommons file, so the file cannot settle a performance per dollar ratio.",false,Intel Newsroom: Intel Arc Pro B-Series GPUs and Xeon 6 Shine in MLPerf Inference v5.1,https://newsroom.intel.com/artificial-intelligence/intel-arc-pro-b-series-gpus-and-xeon-6-shine-in-mlperf-inference-v5-1,2026-10-28,MLPerf Inference v5.1
+N2,Intel,Intel v5.1,Intel Xeon 6 with P-cores achieved 1.9x performance improvement gen-over-gen in MLPerf Inference v5.1,true,Xeon 6 entries in v5.1 against the previous Xeon generation,1.9x generation over generation,"the previous generation is not in the v5.0, v5.1 or v6.0 files",CANNOT_TELL,The earlier-generation Xeon results are in older rounds that these files do not cover.,false,Intel Newsroom: Intel Arc Pro B-Series GPUs and Xeon 6 Shine in MLPerf Inference v5.1,https://newsroom.intel.com/artificial-intelligence/intel-arc-pro-b-series-gpus-and-xeon-6-shine-in-mlperf-inference-v5-1,2026-10-28,MLPerf Inference v5.1
+N3,Intel,Intel v6.0,the Arc Pro B70 providing up to 1.8x higher inference performance than the Arc Pro B60,true,"entries 6.0-0101 (4 x Arc Pro B70) over 6.0-0059 (4 x Arc Pro B60), same host CPU, every closed model and scenario they both ran",up to 1.8x,"ratios gpt-oss-120b offline 1.83x, gpt-oss-120b server 2.10x, llama2-70b-99 offline 1.45x, llama2-70b-99 server 1.54x, llama3.1-8b offline 1.38x, llama3.1-8b server 1.33x, whisper offline 1.40x. Largest offline ratio 1.83x (gpt-oss-120b). Largest of all 2.10x (gpt-oss-120b server)",ABOVE,"The statement says up to. The largest offline ratio matches it and the largest server ratio is higher. The post names the B60 comparison system as 2 Dual GPU cards, 4 B60 equivalent.",false,"Intel Newsroom: Intel Delivers Open, Scalable AI Performance in MLPerf Inference v6.0",https://newsroom.intel.com/artificial-intelligence/intel-delivers-ai-performance-mlperf-inference-v6-0,2026-10-28,MLPerf Inference v6.0
+N6,Intel,Intel v6.0,with over half of MLPerf 6.0 submissions powered by Xeon,true,"the Processor column of the v6.0 file, entries and rows with Xeon in the processor name",over half,"50 of 112 entries (44.6%), 235 of 520 rows (45.2%), 45 of 91 closed datacenter entries (49.5%)",BELOW,"Every count of the processor field in the file is under half. If powered by means something other than the processor field, the file cannot show it.",false,"Intel Newsroom: Intel Delivers Open, Scalable AI Performance in MLPerf Inference v6.0",https://newsroom.intel.com/artificial-intelligence/intel-delivers-ai-performance-mlperf-inference-v6-0,2026-10-28,MLPerf Inference v6.0
+N4,Intel,Intel v6.0,up to 1.18x higher gains on the same Intel Arc Pro B60 hardware versus MLPerf v5.1,true,"entries 6.0-0060 over 5.1-0056, 4 x Arc Pro B60 Dual GPU cards each, every model and scenario both ran",up to 1.18x,"ratios llama2-70b-99 offline 1.087x, llama2-70b-99 server 1.042x, llama3.1-8b offline 1.045x, llama3.1-8b server 1.177x. Largest 1.177x",CONSISTENT,"The host CPU differs, a Xeon w9-3475X in v5.1 and a Xeon 698X in v6.0.",false,"Intel Newsroom: Intel Delivers Open, Scalable AI Performance in MLPerf Inference v6.0",https://newsroom.intel.com/artificial-intelligence/intel-delivers-ai-performance-mlperf-inference-v6-0,2026-10-28,MLPerf Inference v6.0
+N5,Intel,Intel v6.0,powering up to 1.6x as much KV cache capacity when running larger models,true,a memory capacity comparison with competing GPUs,up to 1.6x KV cache capacity,not an MLPerf measurement,CANNOT_TELL,"The MLCommons files hold throughput results, not cache capacity.",false,"Intel Newsroom: Intel Delivers Open, Scalable AI Performance in MLPerf Inference v6.0",https://newsroom.intel.com/artificial-intelligence/intel-delivers-ai-performance-mlperf-inference-v6-0,2026-10-28,MLPerf Inference v6.0
+N7,Intel,Intel v6.0,Intel Xeon 6 processors with P-cores delivered up to a 1.9x generational performance gain in MLPerf Inference v5.1,true,Xeon 6 entries against the previous Xeon generation,up to 1.9x generational gain,"the previous generation is not in the v5.0, v5.1 or v6.0 files",CANNOT_TELL,Same statement as the v5.1 post.,false,"Intel Newsroom: Intel Delivers Open, Scalable AI Performance in MLPerf Inference v6.0",https://newsroom.intel.com/artificial-intelligence/intel-delivers-ai-performance-mlperf-inference-v6-0,2026-10-28,MLPerf Inference v6.0
+P1,Lambda,Lambda v6.0,up to 29% higher iso-GPU throughput than NVIDIA HGX B200 on GPT-OSS-120B,true,"entry 6.0-0063 (4 x GB300) over half of the best closed 8 x B200 results, 6.0-0091 offline and 6.0-0083 server, gpt-oss-120b",up to 29%,"+29.4% offline (60,220.4 vs 46,535.3 tokens/s), +22.3% server (53,462.7 vs 43,722.1)",CONSISTENT,"The post says no 4-GPU B200 entry exists, so the baseline is an 8-GPU result halved. That estimate is the post's own.",false,"Lambda: Lambda's MLPerf Inference v6.0: hardware leap, software maturity, research breakthrough",https://lambda.ai/blog/lambdas-mlperf-inference-v6.0-hardware-leap-software-maturity-research-breakthrough,2026-10-28,MLPerf Inference v6.0
+P2,Lambda,Lambda v6.0,"60,220 tokens/s (Offline) and 53,463 tokens/s (Server)",true,"entry 6.0-0063, gpt-oss-120b","60,220 offline and 53,463 server tokens/s","60,220.4 offline and 53,462.7 server tokens/s",CONSISTENT,matches to rounding,false,"Lambda: Lambda's MLPerf Inference v6.0: hardware leap, software maturity, research breakthrough",https://lambda.ai/blog/lambdas-mlperf-inference-v6.0-hardware-leap-software-maturity-research-breakthrough,2026-10-28,MLPerf Inference v6.0
+P3,Lambda,Lambda v6.0,"130,008 tokens/s (Server), 160,403 tokens/s (Offline), and 128,750 tokens/s (Interactive)",true,"entry 6.0-0062, Llama 3.1 8B","130,008 server, 160,403 offline, 128,750 interactive tokens/s","130,008 server, 160,403 offline, 128,750 interactive",CONSISTENT,matches,false,"Lambda: Lambda's MLPerf Inference v6.0: hardware leap, software maturity, research breakthrough",https://lambda.ai/blog/lambdas-mlperf-inference-v6.0-hardware-leap-software-maturity-research-breakthrough,2026-10-28,MLPerf Inference v6.0
+P4,Lambda,Lambda v6.0,up to a 9% throughput gain on Llama 3.1 8B,true,entry 6.0-0062 over the best v5.1 closed 8 x B200 Llama 3.1 8B result in each scenario,up to 9%,"+9.1% offline over 5.1-0062 (Lenovo), +0.9% server over 5.1-0028 (Dell)",CONSISTENT,The post attributes the gain to a software update from CUDA 12.9 to 13.1. The file lists CUDA 13.1 for 6.0-0062 and CUDA 12.8 for the v5.1 offline entry.,false,"Lambda: Lambda's MLPerf Inference v6.0: hardware leap, software maturity, research breakthrough",https://lambda.ai/blog/lambdas-mlperf-inference-v6.0-hardware-leap-software-maturity-research-breakthrough,2026-10-28,MLPerf Inference v6.0
+P8,Lambda,Lambda v6.0,Dual Intel Xeon Platinum 8570 (56 cores),true,entry 6.0-0062 processor columns,"2 processors, 56 cores","Intel(R) Xeon(R) Platinum 8570, 2 per node, 56 cores",CONSISTENT,Core count is per processor.,false,"Lambda: Lambda's MLPerf Inference v6.0: hardware leap, software maturity, research breakthrough",https://lambda.ai/blog/lambdas-mlperf-inference-v6.0-hardware-leap-software-maturity-research-breakthrough,2026-10-28,MLPerf Inference v6.0
+P5,Lambda,Lambda v6.0,reduces time-to-first-token (TTFT) P99 latency by 31% on GPT-OSS-120B,true,"entry 6.0-0110 (open division), gpt-oss-120b server","TTFT P99 down 31%, throughput 16,167 tokens/s","the file holds throughput only: 16,167.4 tokens/s",CANNOT_TELL,Latency and the baseline run without BLAZE are not in the summary file. The post says these results are not verified by MLCommons Association.,false,"Lambda: Lambda's MLPerf Inference v6.0: hardware leap, software maturity, research breakthrough",https://lambda.ai/blog/lambdas-mlperf-inference-v6.0-hardware-leap-software-maturity-research-breakthrough,2026-10-28,MLPerf Inference v6.0
+P6,Lambda,Lambda v6.0,Lambda is the only AI-native cloud to publish official MLPerf results for both inference and training on an NVIDIA Blackwell Ultra platform,true,Blackwell Ultra entries by submitter in the inference and training results,only,not settled by the inference file,CANNOT_TELL,Needs the training results and a definition of AI-native cloud.,false,"Lambda: Lambda's MLPerf Inference v6.0: hardware leap, software maturity, research breakthrough",https://lambda.ai/blog/lambdas-mlperf-inference-v6.0-hardware-leap-software-maturity-research-breakthrough,2026-10-28,MLPerf Inference v6.0
+P7,Lambda,Lambda v6.0,NVIDIA Blackwell Ultra has 279GB of HBM3e per chip,true,the platform name of entry 6.0-0063,279GB,the entry's platform name reads GB300-NVL72_GB300-288GB_aarch64x4_TRT,CANNOT_TELL,The file names the part with 288GB. The post may mean usable capacity.,false,"Lambda: Lambda's MLPerf Inference v6.0: hardware leap, software maturity, research breakthrough",https://lambda.ai/blog/lambdas-mlperf-inference-v6.0-hardware-leap-software-maturity-research-breakthrough,2026-10-28,MLPerf Inference v6.0
+E1,Nebius,Nebius v6.0,10 first-place results out of 16 benchmark submissions,true,"every Nebius row against all other rows with the same model, scenario, division, accelerator model and GPU count",10 first-place results of 16,10 first places among 16 rows that have a peer in that class,CONSISTENT,The post does not define first place. This rule reproduces its count exactly.,false,"Nebius: MLPerf Inference v6.0 results, top-tier AI performance on NVIDIA Blackwell and Blackwell Ultra",https://nebius.com/blog/posts/mlperf-inference-v6-0-results,2026-10-28,MLPerf Inference v6.0
+E2,Nebius,Nebius v6.0,five additional use cases where we were the only submitter,true,"Nebius rows with no other row in the same class (same model, scenario, division, accelerator model and GPU count)",5,5 rows,CONSISTENT,Same rule as E1.,false,"Nebius: MLPerf Inference v6.0 results, top-tier AI performance on NVIDIA Blackwell and Blackwell Ultra",https://nebius.com/blog/posts/mlperf-inference-v6-0-results,2026-10-28,MLPerf Inference v6.0
+E3,Nebius,Nebius v6.0,Below is the table with all 21 submissions,true,"Nebius rows in the v6.0 file, entries 6.0-0079 to 6.0-0084",21,21 rows,CONSISTENT,A row is one model and scenario of one entry.,false,"Nebius: MLPerf Inference v6.0 results, top-tier AI performance on NVIDIA Blackwell and Blackwell Ultra",https://nebius.com/blog/posts/mlperf-inference-v6-0-results,2026-10-28,MLPerf Inference v6.0
+E4,Nebius,Nebius v6.0,18 nodes with a total of 72 NVIDIA Blackwell Ultra GPUs,true,entry 6.0-0081,"18 nodes, 72 GPUs","18 nodes, 72 GPUs, NVIDIA GB300",CONSISTENT,matches,false,"Nebius: MLPerf Inference v6.0 results, top-tier AI performance on NVIDIA Blackwell and Blackwell Ultra",https://nebius.com/blog/posts/mlperf-inference-v6-0-results,2026-10-28,MLPerf Inference v6.0
+E5,Nebius,Nebius v6.0,The results show linear scaling performance when moving from smaller configurations to the full-rack GB300 NVL72 system,true,"per GPU result of the 72 GPU entry 6.0-0081 over the smaller Nebius GB300 entries (8 GPUs for DeepSeek R1 in 6.0-0082, 1 GPU for gpt-oss 120B in 6.0-0079)",linear scaling,"per GPU ratio deepseek-r1 offline 98.1%, deepseek-r1 server 99.1%, gpt-oss-120b offline 94.5%, gpt-oss-120b server 101.7%",CONSISTENT,Linear would be 100%. The ratios run from 94.5% to 101.7%.,false,"Nebius: MLPerf Inference v6.0 results, top-tier AI performance on NVIDIA Blackwell and Blackwell Ultra",https://nebius.com/blog/posts/mlperf-inference-v6-0-results,2026-10-28,MLPerf Inference v6.0
+E6,Nebius,Nebius v6.0,both server and offline inference scenarios show consistent performance improvements as we move from HGX B200 to HGX B300 and further to the GB300 NVL72 system,true,"DeepSeek R1, 8 GPU entries 6.0-0083 (B200), 6.0-0084 (B300), 6.0-0082 (GB300), tokens/s",improvement at each step,"Offline 58,582, 69,319, 76,347. Server 51,693, 60,413, 64,510",CONSISTENT,Each step is higher than the one before in both scenarios.,false,"Nebius: MLPerf Inference v6.0 results, top-tier AI performance on NVIDIA Blackwell and Blackwell Ultra",https://nebius.com/blog/posts/mlperf-inference-v6-0-results,2026-10-28,MLPerf Inference v6.0
+E7,Nebius,Nebius v6.0,A similar pattern can be seen in the gpt-oss 120B benchmark when comparing HGX B200 and HGX B300 systems running 8 GPUs,true,"gpt-oss 120B, 8 GPU entries 6.0-0083 (B200) and 6.0-0084 (B300), tokens/s",B300 above B200,"Offline 85,921 then 106,885. Server 87,444 then 100,437",CONSISTENT,Both scenarios rise.,false,"Nebius: MLPerf Inference v6.0 results, top-tier AI performance on NVIDIA Blackwell and Blackwell Ultra",https://nebius.com/blog/posts/mlperf-inference-v6-0-results,2026-10-28,MLPerf Inference v6.0
+U1,Supermicro,Supermicro v6.0,"Supermicro posted the top result on the Whisper benchmark, achieving over 50,562 samples per second",true,entry 6.0-0098 against every closed datacenter Whisper offline result in v6.0 (35 rows),"top result, over 50,562 samples/s","50,562.2 samples/s, rank 1 of 35. Next: Google 49,327.2 (6.0-0040)",CONSISTENT,Whisper has an offline scenario only. The result is the total for one 8-GPU system.,false,Supermicro Data Center Stories: Supermicro Leads Whisper Benchmark in MLPerf v6.0 with NVIDIA Blackwell Ultra GPUs,https://learn-more.supermicro.com/data-center-stories/supermicro-leads-whisper-benchmark-in-mlperf-v6-nvidia-b300-gpus,2026-10-28,MLPerf Inference v6.0
+U2,Supermicro,Supermicro v6.0,dual AMD EPYC™ 9575F CPUs,true,processor columns of entry 6.0-0098,2 x AMD EPYC 9575F,"AMD EPYC 9575F, 2 per node",CONSISTENT,matches,false,Supermicro Data Center Stories: Supermicro Leads Whisper Benchmark in MLPerf v6.0 with NVIDIA Blackwell Ultra GPUs,https://learn-more.supermicro.com/data-center-stories/supermicro-leads-whisper-benchmark-in-mlperf-v6-nvidia-b300-gpus,2026-10-28,MLPerf Inference v6.0
+U3,Supermicro,Supermicro v6.0,2.3 TB of HBM3e GPU memory per system,true,"accelerator and platform names of entry 6.0-0098, 8 GPUs",2.3 TB,"accelerator column NVIDIA B300-SXM-270GB (8 x 270 GB = 2.16 TB), platform name AS-8126GS-NB3RT_B300-SXM-288GB_TRT (8 x 288 GB = 2.30 TB)",CANNOT_TELL,The two names in the entry disagree about the memory per GPU. The stated figure matches the platform name.,false,Supermicro Data Center Stories: Supermicro Leads Whisper Benchmark in MLPerf v6.0 with NVIDIA Blackwell Ultra GPUs,https://learn-more.supermicro.com/data-center-stories/supermicro-leads-whisper-benchmark-in-mlperf-v6-nvidia-b300-gpus,2026-10-28,MLPerf Inference v6.0
+V1,Supermicro,Supermicro Training v6.0,"demonstrated linear scaling in training performance for 3 different models, scaling from 8 GPU to 72 B300 GPUs",true,"Supermicro 8 x B300 and 72 x GB300 result sets, official-style time to train (olympic average with scaling factor), benchmarks with both: Llama 3.1 8B, Llama 2 70B LoRA, GPT-OSS 20B","linear scaling for 3 models, 8 to 72 GPUs","llama31_8b 8 GPUs 74.787 min, 72 GPUs 11.860 min, speedup 6.31x for 9x the GPUs (70% of linear), 16 GPUs 55.050 min, speedup 1.36x for 2x (68% of linear). llama2_70b_lora 8 GPUs 6.674 min, 72 GPUs 1.173 min, speedup 5.69x for 9x the GPUs (63% of linear), 16 GPUs 6.259 min, speedup 1.07x for 2x (53% of linear). gpt_oss_20b 8 GPUs 84.283 min, 72 GPUs 18.186 min, speedup 4.63x for 9x the GPUs (51% of linear)",BELOW,"Linear would cut time to train by 9x. The file shows 51% to 70% of that. The post does not define performance, and larger runs change batch size. Time to train is the primary MLPerf metric.",false,Supermicro Data Center Stories: Supermicro NVIDIA Blackwell Systems Demonstrate Linear Scalability for MLPerf Training v6.0,https://learn-more.supermicro.com/data-center-stories/supermicro-nvidia-blackwell-systems-linear-scalability-mlperf-training,2026-10-28,MLPerf Training v6.0
+V2,Supermicro,Supermicro Training v6.0,2 systems of HGX-B300 showed even better performance than the linear scaling for HGX-B300 to NVL72,true,"Llama 3.1 8B, 16 GPU entry against the straight line (linear axes) and the power-law line (log axes) joining the 8 and 72 GPU entries",16 GPUs above the 8 to 72 GPU scaling line,16 GPUs give a 1.36x speedup over 8 GPUs. The line gives 1.66x (linear axes) or 1.79x (log axes),BELOW,The 16 GPU point sits under the line on both axis choices. The post does not say which axes.,false,Supermicro Data Center Stories: Supermicro NVIDIA Blackwell Systems Demonstrate Linear Scalability for MLPerf Training v6.0,https://learn-more.supermicro.com/data-center-stories/supermicro-nvidia-blackwell-systems-linear-scalability-mlperf-training,2026-10-28,MLPerf Training v6.0
+V3,Supermicro,Supermicro Training v6.0,the 2 systems of HGX-B300 follow the linear scaling of the HGX-B300 to NVL72,true,"Llama 2 70B LoRA, 16 GPU entry against the same two lines",16 GPUs on the 8 to 72 GPU scaling line,16 GPUs give a 1.07x speedup over 8 GPUs. The line gives 1.59x (linear axes) or 1.73x (log axes),BELOW,The post calls the model Llama 3.1 70b. The benchmark in the repository is llama2_70b_lora.,false,Supermicro Data Center Stories: Supermicro NVIDIA Blackwell Systems Demonstrate Linear Scalability for MLPerf Training v6.0,https://learn-more.supermicro.com/data-center-stories/supermicro-nvidia-blackwell-systems-linear-scalability-mlperf-training,2026-10-28,MLPerf Training v6.0
+V4,Supermicro,Supermicro Training v6.0,B300 performs better than B200 in Training performance,true,"Supermicro 8 GPU entries, official-style time to train, B300 against B200",B300 faster than B200,llama31_8b B300 74.79 min against B200 83.68. llama2_70b_lora B300 6.67 min against B200 7.86. gpt_oss_20b B300 84.28 min against B200 98.77. dlrm_dcnv2 B300 2.18 min against B200 2.31,CONSISTENT,B300 is faster on all four benchmarks where Supermicro ran both.,false,Supermicro Data Center Stories: Supermicro NVIDIA Blackwell Systems Demonstrate Linear Scalability for MLPerf Training v6.0,https://learn-more.supermicro.com/data-center-stories/supermicro-nvidia-blackwell-systems-linear-scalability-mlperf-training,2026-10-28,MLPerf Training v6.0
+D9,Dell,Dell v6.1,"The R770AP table in the post lists Whisper Offline at 2,200.80 for Intel v6.0, 2,091.60 for Vendor B v6.0 and 2,275.93 for Dell R770AP v6.1, which it states as +3.4% against Intel and +8.8% against Vendor B.",false,"Dell R770AP entry 6.1-0031 Whisper Offline 2,275.93 against v6.0 Intel 6.0-0057 and Supermicro 6.0-0096","+3.4% and +8.8%, Vendor B v6.0 at 2,091.60","Intel 2,200.80 (+3.41%). Supermicro v6.0-0096 is 2,204.09 (+3.26%). No v6.0 row has 2,091.60",BELOW,The Intel row matches. The Vendor B baseline in the post differs from the Supermicro v6.0 entry on the same CPU.,false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+D1,Dell,Dell v6.1,Llama 3.1-8B Server throughput on the R770 more than double generation-over-generation (+110.1%).,true,"Dell R770 (Xeon 6787P, 2 x 86 cores) Llama 3.1-8B Server, v6.0 entry 6.0-0018 against v6.1 entry 6.1-0032",+110.1%,"v6.0 entry 6.0-0018 245.63, v6.1 entry 6.1-0032 516.01, change +110.1%",CONSISTENT,The ratio is 2.101.,false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+D10,Dell,Dell v6.1,Dell comes out ahead on Whisper Offline (+1.2%),true,"Dell R770AP v6.1 Whisper Offline against Supermicro entry 6.1-0095 (128 cores), Whisper Offline",+1.2%,"2,275.93 against 2,248.15 (+1.24%)",CONSISTENT,,false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+D11,Dell,Dell v6.1,is the only other RGAT submitter on this platform,true,every closed RGAT entry on the Xeon 6980P in v6.1,Dell the only other RGAT submitter besides Vendor B,"Dell 21,835.4, Intel 24,573.7",CONSISTENT,"Two submitters list RGAT on this CPU, Intel and Dell. Supermicro lists none. We read ""other"" as other than Intel.",false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+D13,Dell,Dell v6.1,"The two configuration tables in the post list the R770 as Xeon 6787P with 86 cores per socket and 16 x 64 GB RDIMM at 6400 MT/s, and the R770AP as Xeon 6980P with 128 cores per socket and 24 x 96 GB RDIMM at 6400 MT/s.",false,"processor, cores and system name of the Dell R770 and R770AP v6.1 entries",6787P 86 cores 16x64 GB at 6400. 6980P 128 cores 24x96 GB at 6400,"PowerEdge R770AP_128C 24x96GB (6400 MT/s), Intel(R) Xeon(R) 6980P, 2 x 128 cores. PowerEdge R770_86C 16x64GB (6400 MT/s), Intel(R) Xeon(R) 6787P, 2 x 86 cores",CONSISTENT,"All six entries are closed division, datacenter and available.",false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+D2,Dell,Dell v6.1,"The generational table in the post lists R770 Llama 3.1-8B Offline at 807.7 in v6.0 and 1,228.0 in v6.1, an uplift of +52.0%.",false,"same two entries, Llama 3.1-8B Offline",+52.0%,"v6.0 entry 6.0-0018 807.67, v6.1 entry 6.1-0032 1,227.99, change +52.0%",CONSISTENT,,false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+D3,Dell,Dell v6.1,"The generational table in the post lists R770 RGAT Offline at 14,283.7 in v6.0 and 16,042.8 in v6.1, an uplift of +12.3%.",false,"same two entries, RGAT Offline",+12.3%,"v6.0 entry 6.0-0018 14,283.70, v6.1 entry 6.1-0032 16,042.80, change +12.3%",CONSISTENT,,false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+D4,Dell,Dell v6.1,"The generational table in the post lists R770 Whisper Offline at 1,414.3 in v6.0 and 1,539.5 in v6.1, an uplift of +8.9%.",false,"same two entries, Whisper Offline",+8.9%,"v6.0 entry 6.0-0018 1,414.26, v6.1 entry 6.1-0032 1,539.49, change +8.9%",CONSISTENT,The post labels the Whisper unit tokens/sec. The files list samples/s.,false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+D5,Dell,Dell v6.1,"The R770 outperformed a same-CPU competitive system on 3 of 4 workloads, by as much as +20.8% on RGAT.",true,"Dell R770 entries against the Supermicro entry 6.1-0096 (1-node-2S-GNR_86C), which is the only other Xeon 6787P entry in the v6.1 file. The post does not name Vendor B","3 of 4 workloads, up to +20.8%","Supermicro against Dell: llama3_1-8b Server: 516.06 against 516.01 (-0.01%). llama3_1-8b Offline: 1,189.89 against 1,227.99 (+3.20%). rgat Offline: 13,273.70 against 16,042.80 (+20.86%). Whisper Offline: 1,477.40 against 1,539.49 (+4.20%)",CONSISTENT,"Dell leads on 3 of 4. RGAT is +20.86%, which the post rounds down to +20.8%. We take the competitor to be the Supermicro entry because its four values match the Vendor B column of the post.",false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+D8,Dell,Dell v6.1,"The R770AP table in the post lists RGAT Offline at 24,149.80 for Intel v6.0, 25,192.70 for Vendor B v6.0 and 21,835.40 for Dell R770AP v6.1, which it states as -9.5% against Intel and -13.3% against Vendor B.",false,"Dell R770AP entry 6.1-0031 RGAT Offline 21,835.4 against v6.0 Intel 6.0-0057 and Supermicro 6.0-0096 (same CPU, 128 cores)",-9.5% and -13.3%,"21,835.4 against Intel 24,149.8 (-9.58%) and Supermicro 25,192.7 (-13.33%)",CONSISTENT,The Intel change rounds to -9.6% against the stated -9.5%.,false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+D12,Dell,Dell v6.1,"Vendor B's configuration used 24x96 GB MRDIMMs running at 8,800 MT/s, versus the R770AP's 24x96 GB RDIMMs at 6,400 MT/s",true,memory fields of the Supermicro 128 core v6.1 entries and the Dell R770AP entry,"MRDIMM 8,800 MT/s against RDIMM 6,400 MT/s",The Dell entry name reads 24x96GB (6400 MT/s). The Supermicro entry name and notes carry no memory description,CANNOT_TELL,The Supermicro memory type and speed are not in the summary file. The statement that Intel confirmed the cause is not in any file.,false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+D6,Dell,Dell v6.1,The R770 server leads on 3 of 4 workloads with zero performance losses.,true,Llama 3.1-8B Server in the same two entries,"zero losses, Llama Server a statistical tie","Dell 516.012 against 516.059, Dell lower by 0.009%",CANNOT_TELL,"The Dell value is lower by a few thousandths of a percent. The files hold one result per entry and no spread between runs, so a tie cannot be tested.",false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+D7,Dell,Dell v6.1,The R770AP's debut submission posted a +38.7% Llama Offline gain over the Intel reference baseline.,true,"Dell R770AP entry 6.1-0031 in the v6.1 file, Llama 3.1-8B. v6.0 Intel entry 6.0-0057 Llama Offline 1,229.56","+38.7%, 1,706.45 tokens/s","No Dell R770AP Llama 3.1-8B entry in the v6.1 file (0 rows). The file lists Dell R770AP for RGAT and Whisper only. Intel on the same CPU is listed at 1,916.74 in v6.1",CANNOT_TELL,"The post says elsewhere that it did not publish Llama on the R770AP. The 1,706.45 value is not in the file.",false,"Dell Technologies: Faster Inference, Fewer Compromises: MLPerf v6.1 Results on the Dell PowerEdge R770 and R770AP (post dated 23 October 2026)",https://www.dell.com/en-us/infohub/articles/faster-inference-fewer-compromises-mlperf-v6-1-results-on-the-dell-poweredge-r770-and-r770ap,2026-10-28,MLPerf Inference v6.1
+M1,Dell,Dell MI355X v6.0,"achieving 70,610.8 tokens/s",true,"entry 6.0-0020 (Dell XE9785L, 8 x MI355X), Llama 2 70B Interactive","70,610.8 tokens/s","70,610.8 tokens/s",CONSISTENT,,false,Dell Technologies Info Hub: MLPerf Inference v6.0: Dell Showcases Breakthrough Performance with AMD Instinct MI355X GPUs (post dated 16 April 2026),https://infohub.delltechnologies.com/en-us/p/mlperf-inference-v6-0-dell-showcases-breakthrough-performance-with-amd-instinct-tm-mi355x-gpus/,2026-10-28,MLPerf Inference v6.0
+M2,Dell,Dell MI355X v6.0,The MI355X delivers nearly 4x higher inference performance compared to the MI300X.,true,"entry 6.0-0020 (8 x MI355X) against entry 6.0-0019 (Dell XE9680, 8 x MI300X), Llama 2 70B",nearly 4x,"Offline 3.69x, Server 3.95x. Entry 6.0-0019 has no Interactive row",CONSISTENT,The post says three scenarios. The file has two for the MI300X entry.,false,Dell Technologies Info Hub: MLPerf Inference v6.0: Dell Showcases Breakthrough Performance with AMD Instinct MI355X GPUs (post dated 16 April 2026),https://infohub.delltechnologies.com/en-us/p/mlperf-inference-v6-0-dell-showcases-breakthrough-performance-with-amd-instinct-tm-mi355x-gpus/,2026-10-28,MLPerf Inference v6.0
+M3,Dell,Dell MI355X v6.0,"Table 1 in the post lists GPT-OSS-120B on the XE9785L with 8 x MI355X at 78,603.5 queries/s in Server and 91,395.5 samples/s in Offline.",false,entry 6.0-0020 GPT-OSS-120B,"Server 78,603.5, Offline 91,395.5","Server 78,603.5 tokens/s, Offline 91,395.5 tokens/s",CONSISTENT,The values match. The post labels the units queries/s and samples/s. The file lists tokens/s for both.,false,Dell Technologies Info Hub: MLPerf Inference v6.0: Dell Showcases Breakthrough Performance with AMD Instinct MI355X GPUs (post dated 16 April 2026),https://infohub.delltechnologies.com/en-us/p/mlperf-inference-v6-0-dell-showcases-breakthrough-performance-with-amd-instinct-tm-mi355x-gpus/,2026-10-28,MLPerf Inference v6.0
+M5,Dell,Dell MI355X v6.0,"Approximately 95.5% scaling efficiency was achieved, despite mixing GPU generations across continents.",true,"entry 6.0-0031 (Dell MangoBoost heterogeneous cluster, 24 GPUs) against the sum of single-system entries 6.0-0069 (8 x MI300X), 6.0-0070 (8 x MI325X) and 6.0-0032 (8 x MI355X), Llama 2 70B",about 95.5%,"Offline 151,843 against a sum of 159,014 = 95.5%. Server 141,521 against 150,624 = 94.0%",CONSISTENT,The post does not name its baseline. The Offline figure reproduces with these three entries. Server is lower. The regions are not in the file.,false,Dell Technologies Info Hub: MLPerf Inference v6.0: Dell Showcases Breakthrough Performance with AMD Instinct MI355X GPUs (post dated 16 April 2026),https://infohub.delltechnologies.com/en-us/p/mlperf-inference-v6-0-dell-showcases-breakthrough-performance-with-amd-instinct-tm-mi355x-gpus/,2026-10-28,MLPerf Inference v6.0
+M6,Dell,Dell MI355X v6.0,Only ~16-18% throughput reduction for a 29% power drop,true,"entry 6.0-0033 (XE9785, MI355X power capped at 1000 W) against entry 6.0-0032 (same system, 1400 W), Llama 2 70B",16% to 18% lower,"Offline -16.1%, Server -18.1%. 1,000 W against 1,400 W is -28.6%",CONSISTENT,,false,Dell Technologies Info Hub: MLPerf Inference v6.0: Dell Showcases Breakthrough Performance with AMD Instinct MI355X GPUs (post dated 16 April 2026),https://infohub.delltechnologies.com/en-us/p/mlperf-inference-v6-0-dell-showcases-breakthrough-performance-with-amd-instinct-tm-mi355x-gpus/,2026-10-28,MLPerf Inference v6.0
+M4,Dell,Dell MI355X v6.0,The XE9785L (8x MI355X) also delivered the first MLPerf submission of the open-source GPT-OSS-120B model,true,closed GPT-OSS-120B entries in v6.0,first submission of GPT-OSS-120B,"14 submitters list GPT-OSS-120B in v6.0: AMD, ASUSTeK, Cisco, CoreWeave, Dell, GigaComputing, HPE, Intel, Lambda_SIT, Lenovo, NVIDIA, Nebius, ORACLE, RedHat",CANNOT_TELL,v6.0 is the first round with this model. The file does not order submissions within the round.,false,Dell Technologies Info Hub: MLPerf Inference v6.0: Dell Showcases Breakthrough Performance with AMD Instinct MI355X GPUs (post dated 16 April 2026),https://infohub.delltechnologies.com/en-us/p/mlperf-inference-v6-0-dell-showcases-breakthrough-performance-with-amd-instinct-tm-mi355x-gpus/,2026-10-28,MLPerf Inference v6.0
+M7,Dell,Dell MI355X v6.0,17.5% better tokens/s per Watt efficiency at the 1000W cap,true,"same two entries, throughput divided by the capped power (1,000 W against 1,400 W)",+17.5%,"Offline +17.5%, Server +14.6% using the cap values. Neither entry has measured power",CANNOT_TELL,The cap values are not measured power. The Offline figure reproduces from them and the Server figure does not.,false,Dell Technologies Info Hub: MLPerf Inference v6.0: Dell Showcases Breakthrough Performance with AMD Instinct MI355X GPUs (post dated 16 April 2026),https://infohub.delltechnologies.com/en-us/p/mlperf-inference-v6-0-dell-showcases-breakthrough-performance-with-amd-instinct-tm-mi355x-gpus/,2026-10-28,MLPerf Inference v6.0
+O1,Dell,Dell Overview v6.1,"The XE9785L produced 103,932 tokens/s on Llama2 70B at 99.9% Server and 110,188 tokens/s on GPT-OSS-120B Server",true,"v6.1 Dell XE9785L_8xMI355X, Llama 2 70B 99.9 Server and GPT-OSS-120B Server","103,932 and 110,188","103,932 and 110,188",CONSISTENT,,false,Dell Technologies Info Hub: Driving AI Innovation: Dell Technologies Advances AI Inference Performance with MLPerf Inference v6.1 (post dated 16 September 2026),https://infohub.delltechnologies.com/en-uk/p/driving-ai-innovation-dell-technologies-advances-ai-inference-performance-with-mlperf-inference-v6-1/,2026-10-28,MLPerf Inference v6.1
+O10,Dell,Dell Overview v6.1,"The configuration delivered 253,501 tokens/s on GPT-OSS-120B Server and 285,454 tokens/s Offline",true,"same entry, GPT-OSS-120B","253,501 and 285,454","253,501 and 285,454",CONSISTENT,,false,Dell Technologies Info Hub: Driving AI Innovation: Dell Technologies Advances AI Inference Performance with MLPerf Inference v6.1 (post dated 16 September 2026),https://infohub.delltechnologies.com/en-uk/p/driving-ai-innovation-dell-technologies-advances-ai-inference-performance-with-mlperf-inference-v6-1/,2026-10-28,MLPerf Inference v6.1
+O2,Dell,Dell Overview v6.1,As the only MLPerf Inference v6.1 submission featuring the AMD Instinct(TM) MI350P,true,every v6.1 row whose accelerator is MI350P,only submission with MI350P,Submitters with MI350P: Dell_AMD,CONSISTENT,The submitter name in the file is Dell_AMD.,false,Dell Technologies Info Hub: Driving AI Innovation: Dell Technologies Advances AI Inference Performance with MLPerf Inference v6.1 (post dated 16 September 2026),https://infohub.delltechnologies.com/en-uk/p/driving-ai-innovation-dell-technologies-advances-ai-inference-performance-with-mlperf-inference-v6-1/,2026-10-28,MLPerf Inference v6.1
+O3,Dell,Dell Overview v6.1,Dell demonstrated the massive scale of NVDIA GB300 infrastructure with 72 GB300 288GB GPUs spanning 18 PowerEdge XE9712 nodes.,true,v6.1 Dell PowerEdge XE9712 entry,"72 GPUs, 18 nodes","72 accelerators, 18 nodes, NVIDIA GB300",CONSISTENT,,false,Dell Technologies Info Hub: Driving AI Innovation: Dell Technologies Advances AI Inference Performance with MLPerf Inference v6.1 (post dated 16 September 2026),https://infohub.delltechnologies.com/en-uk/p/driving-ai-innovation-dell-technologies-advances-ai-inference-performance-with-mlperf-inference-v6-1/,2026-10-28,MLPerf Inference v6.1
+O4,Dell,Dell Overview v6.1,delivering 10.585 seconds of SingleStream latency and 0.646 seconds Offline performance on WAN 2.2 T2V A14B.,true,v6.1 Dell XE9712 WAN 2.2 SingleStream and Offline,10.585 s and 0.646 s,SingleStream 10.585 s. Offline 0.6455 samples/s,CONSISTENT,Both values match. The file lists Offline in samples per second. The post writes seconds.,false,Dell Technologies Info Hub: Driving AI Innovation: Dell Technologies Advances AI Inference Performance with MLPerf Inference v6.1 (post dated 16 September 2026),https://infohub.delltechnologies.com/en-uk/p/driving-ai-innovation-dell-technologies-advances-ai-inference-performance-with-mlperf-inference-v6-1/,2026-10-28,MLPerf Inference v6.1
+O5,Dell,Dell Overview v6.1,"the XE9785L B300 configuration achieved 113,101 tokens/s on Llama 2 70B Offline",true,v6.1 Dell XE9785L (8 x B300) Llama 2 70B 99.9 Offline,"113,101","113,101",CONSISTENT,,false,Dell Technologies Info Hub: Driving AI Innovation: Dell Technologies Advances AI Inference Performance with MLPerf Inference v6.1 (post dated 16 September 2026),https://infohub.delltechnologies.com/en-uk/p/driving-ai-innovation-dell-technologies-advances-ai-inference-performance-with-mlperf-inference-v6-1/,2026-10-28,MLPerf Inference v6.1
+O7,Dell,Dell Overview v6.1,"reaching 28,973 tokens/s in Server and 29,797 tokens/s in Offline",true,v6.1 Dell XE7740 (8 x RTX PRO 6000) Llama 2 70B 99.9,"28,973 and 29,797","28,973.3 and 29,796.7",CONSISTENT,,false,Dell Technologies Info Hub: Driving AI Innovation: Dell Technologies Advances AI Inference Performance with MLPerf Inference v6.1 (post dated 16 September 2026),https://infohub.delltechnologies.com/en-uk/p/driving-ai-innovation-dell-technologies-advances-ai-inference-performance-with-mlperf-inference-v6-1/,2026-10-28,MLPerf Inference v6.1
+O8,Dell,Dell Overview v6.1,highlighting the 13.5% and 10.2% improvement in the Server and Offline scenarios respectively for the Llama 2 70B model,true,"Dell XE7740 v6.0 entry 6.0-0022 against the v6.1 entry, Llama 2 70B 99.9","+13.5% Server, +10.2% Offline","Server +13.5%, Offline +10.2%",CONSISTENT,The v6.0 XE7745 entry 6.0-0023 gives +13.5% Server and +11.1% Offline.,false,Dell Technologies Info Hub: Driving AI Innovation: Dell Technologies Advances AI Inference Performance with MLPerf Inference v6.1 (post dated 16 September 2026),https://infohub.delltechnologies.com/en-uk/p/driving-ai-innovation-dell-technologies-advances-ai-inference-performance-with-mlperf-inference-v6-1/,2026-10-28,MLPerf Inference v6.1
+O9,Dell,Dell Overview v6.1,Dell's v6.1 submission demonstrates this concept with a 32-GPU heterogeneous four-cluster configuration combining AMD Instinct(TM) MI300X and MI355X accelerators.,true,v6.1 Dell_MangoBoost heterogeneous entry,"32 GPUs, MI300X and MI355X","32 accelerators, system name Heterogeneous Four-Cluster (2x8x MI300X + 2x8x MI355X, LLMBoost)",CONSISTENT,,false,Dell Technologies Info Hub: Driving AI Innovation: Dell Technologies Advances AI Inference Performance with MLPerf Inference v6.1 (post dated 16 September 2026),https://infohub.delltechnologies.com/en-uk/p/driving-ai-innovation-dell-technologies-advances-ai-inference-performance-with-mlperf-inference-v6-1/,2026-10-28,MLPerf Inference v6.1
+O6,Dell,Dell Overview v6.1,"70,494 tokens/s on DeepSeek-R1 Offline",true,every v6.1 DeepSeek-R1 row from a Dell submitter,"70,494 tokens/s Offline","0 DeepSeek-R1 rows from Dell, Dell_AMD or Dell_MangoBoost in the v6.1 file",CANNOT_TELL,The file has no Dell DeepSeek-R1 row to compare with.,false,Dell Technologies Info Hub: Driving AI Innovation: Dell Technologies Advances AI Inference Performance with MLPerf Inference v6.1 (post dated 16 September 2026),https://infohub.delltechnologies.com/en-uk/p/driving-ai-innovation-dell-technologies-advances-ai-inference-performance-with-mlperf-inference-v6-1/,2026-10-28,MLPerf Inference v6.1
+Y1,Dell,Dell YOLO v6.0,The published v6.0 result for the Dell Pro Max with GB10 on yolo-95 and yolo-99 Offline shows 1.75 samples per second.,true,"entry 6.0-0028, YOLO Offline, both models",1.75 samples/s,"yolo-95 1.74887, yolo-99 1.74887 samples/s",CONSISTENT,,false,Dell Technologies Info Hub: Benchmark Builders to Performance Breakers: Dell's YOLO Journey in MLPerf Inference v6.0 (post dated 24 April 2026),https://infohub.delltechnologies.com/en-us/p/benchmark-builders-to-performance-breakers-dell-s-yolo-journey-in-mlperf-inference-v6-0/,2026-10-28,MLPerf Inference v6.0
+Y2,Dell,Dell YOLO v6.0,Dell was one of two organizations to submit YOLO results for the edge suite: the other was GATEOverflow,true,closed edge YOLO entries in v6.0,two organizations,"Submitters: Dell, GATEOverflow",CONSISTENT,,false,Dell Technologies Info Hub: Benchmark Builders to Performance Breakers: Dell's YOLO Journey in MLPerf Inference v6.0 (post dated 24 April 2026),https://infohub.delltechnologies.com/en-us/p/benchmark-builders-to-performance-breakers-dell-s-yolo-journey-in-mlperf-inference-v6-0/,2026-10-28,MLPerf Inference v6.0
+Y3,Dell,Dell YOLO v6.0,"Table 1 in the post lists published YOLO results for five systems as Offline, SingleStream and MultiStream: Dell Pro Max (CPU only) 1.75, 611.93 and 4,891.38, Dell PowerEdge XE7745 with 16 L4 102.23, 10.55 and 81.09, Dell PowerEdge XE7745 with 8 L40S 153.54, 7.12 and 54.67, GATEOverflow Intel i9 12.07, 90.11 and 843.37, GATEOverflow GTX 1650 Ti 24.19, 43.15 and 344.44.",false,"entries 6.0-0028, 6.0-0030, 6.0-0029, 6.0-0036 and 6.0-0034, YOLO Offline, SingleStream and MultiStream",fifteen published values,"6.0-0028: 1.75, 611.93, 4891.38. 6.0-0030: 102.23, 10.55, 81.09. 6.0-0029: 153.54, 7.12, 54.67. 6.0-0036: 12.07, 90.11, 843.37. 6.0-0034: 24.19, 43.15, 344.44",CONSISTENT,The corrected GB10 row in the same table is not in the file.,false,Dell Technologies Info Hub: Benchmark Builders to Performance Breakers: Dell's YOLO Journey in MLPerf Inference v6.0 (post dated 24 April 2026),https://infohub.delltechnologies.com/en-us/p/benchmark-builders-to-performance-breakers-dell-s-yolo-journey-in-mlperf-inference-v6-0/,2026-10-28,MLPerf Inference v6.0
+Y4,Dell,Dell YOLO v6.0,roughly 10x ahead of the L4 at 6.39 samples per second and roughly 3.4x ahead of the L40S at 19.19 samples per second,true,"entries 6.0-0030 (16 x L4) and 6.0-0029 (8 x L40S), YOLO Offline divided by the GPU count",6.39 and 19.19 per GPU,6.39 and 19.19 per GPU,CANNOT_TELL,"The ratios are the part of the sentence the file cannot settle. They use the corrected GB10 value, which is not in the file.",false,Dell Technologies Info Hub: Benchmark Builders to Performance Breakers: Dell's YOLO Journey in MLPerf Inference v6.0 (post dated 24 April 2026),https://infohub.delltechnologies.com/en-us/p/benchmark-builders-to-performance-breakers-dell-s-yolo-journey-in-mlperf-inference-v6-0/,2026-10-28,MLPerf Inference v6.0
+Y5,Dell,Dell YOLO v6.0,the corrected Offline throughput for the Dell Pro Max with GB10 on both yolo-95 and yolo-99 is: 64.07028 samples per second. This is a ~37x improvement over the published CPU result of 1.75 samples per second,true,"Dell Pro Max GB10 entries in the v6.0 file, and the later v6.1 file","64.07028 samples/s, about 37x",v6.0 file: 1.75 samples/s. The post gives the corrected run as not official. The v6.1 file lists Dell Pro Max (1x GB10) YOLO Offline at 78.7257 samples/s,CANNOT_TELL,The post prints 64.07028 and also 64.7028. Over 1.75 they give 36.6x and 37.0x. The corrected run is not in any MLCommons file.,false,Dell Technologies Info Hub: Benchmark Builders to Performance Breakers: Dell's YOLO Journey in MLPerf Inference v6.0 (post dated 24 April 2026),https://infohub.delltechnologies.com/en-us/p/benchmark-builders-to-performance-breakers-dell-s-yolo-journey-in-mlperf-inference-v6-0/,2026-10-28,MLPerf Inference v6.0
+Y6,Dell,Dell YOLO v6.0,"The table above shows a significant and expected drop in latency when running on the GPU, nearly 51x in the SingleStream and MultiStream scenarios.",true,published GB10 latencies (entry 6.0-0028) over the corrected latencies in the post table (12.02 ms and 94.52 ms),nearly 51x,Published 611.93 ms and 4891.38 ms. Over the corrected values: 50.9x and 51.7x. The v6.1 file lists 13.87 ms and 105.16 ms,CANNOT_TELL,The corrected latencies are not official numbers.,false,Dell Technologies Info Hub: Benchmark Builders to Performance Breakers: Dell's YOLO Journey in MLPerf Inference v6.0 (post dated 24 April 2026),https://infohub.delltechnologies.com/en-us/p/benchmark-builders-to-performance-breakers-dell-s-yolo-journey-in-mlperf-inference-v6-0/,2026-10-28,MLPerf Inference v6.0
+Y7,Dell,Dell YOLO v6.0,Table 1 in the post lists the GATEOverflow GTX 1650 Ti system at a spec power of 5 W and an Offline efficiency of 0.484 samples per second per watt.,false,"spec power and efficiency columns of the table, entry 6.0-0034 (GTX 1650 Ti) Offline 24.187",5 W and 0.484 samples/s/W,The v6.0 file has no power for these entries. 24.187 over 5 W is 4.84. 24.187 over 0.484 is 50 W. The post text gives the card as 80 W,CANNOT_TELL,The efficiency column for this row does not follow from the 5 W in the same row.,false,Dell Technologies Info Hub: Benchmark Builders to Performance Breakers: Dell's YOLO Journey in MLPerf Inference v6.0 (post dated 24 April 2026),https://infohub.delltechnologies.com/en-us/p/benchmark-builders-to-performance-breakers-dell-s-yolo-journey-in-mlperf-inference-v6-0/,2026-10-28,MLPerf Inference v6.0
+H3,HPE,HPE v6.0,Our flagship HPE ProLiant Compute DL380a Gen12 claimed eight No. 1 positions in Llama-focused tests,true,"Llama rows of entry 6.0-0047 (8 x RTX PRO 6000) in the v6.0 file, ranked within the same accelerator and count",eight No. 1 positions in Llama tests,Entry 6.0-0047 has 7 Llama rows. First in group: 7 of 7. Including Whisper: 8,BELOW,"The file has no Llama 2 70B Interactive row for this entry. At most seven Llama rows exist, so eight cannot be reached.",false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H10,HPE,HPE v6.0,"outperforming four competitive systems with an outstanding 4712 samples per second, yielding 2356 samples per second per GPU on a two-GPU configuration",true,"entry 6.0-0052 Whisper Offline, per GPU, against other closed datacenter Whisper Offline entries with the same accelerator","4,712 total, 2,356 per GPU, ahead of four systems","4,712.66 total, 2,356.3 per GPU on 2 GPUs. Non-HPE entries with the same accelerator and a lower per-GPU result: 7",CONSISTENT,The post does not say which four. The file shows more than four.,false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H11,HPE,HPE v6.0,"Table 1 in the post lists HPE DL385 Gen11 against Cisco UCS X215c M8 in Llama 3.1-8B Server and Offline, Llama 2 70B 99 Server and Offline, and Whisper Offline, with ten values.",false,"entries 6.0-0054 (HPE DL385, 4 x RTX PRO 6000) and 6.0-0014 (Cisco UCS X215c M8, 4 x RTX PRO 6000), ten values",Table 1 values,"llama3.1-8b Server: HPE 23,227.2, Cisco 23,423.9. llama3.1-8b Offline: HPE 23,556.7, Cisco 24,013.5. llama2-70b-99 Server: HPE 12,822.2, Cisco 13,029.4. llama2-70b-99 Offline: HPE 13,361.9, Cisco 13,529.8. Whisper Offline: HPE 9,228.4, Cisco 9,099.1",CONSISTENT,All ten values reproduce. Cisco is higher in the four Llama rows by 0.8% to 2.0%. HPE is higher in Whisper by 1.4%.,false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H13,HPE,HPE v6.0,HPE was the only vendor in MLPerf® v6.0 to submit systems using the NVIDIA RTX PRO 4500 Blackwell Server Edition GPU,true,every v6.0 row whose accelerator is RTX PRO 4500,only vendor with RTX PRO 4500,"Entries with that accelerator: HPE 6.0-0043, HPE 6.0-0050, HPE 6.0-0055, HPE 6.0-0056",CONSISTENT,Footnote 15 cites 6.0-0043 (DL145) and 6.0-0050 (DL380 Gen12). Neither is a DL385.,false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H14,HPE,HPE v6.0,"In MLPerf® Inference v5.1, the DL385 Gen11 secured the No. 1 position in the Whisper speech-to-text benchmark",true,"entry 5.1-0052, v5.1 closed datacenter Whisper Offline","No. 1 in Whisper, v5.1","7,925.6 samples/s. Rank 1 of 1 among entries with the same accelerator and count. Highest overall: NVIDIA 45,333.4",CONSISTENT,Rank across all entries is not first. The post gives no group.,false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H4,HPE,HPE v6.0,outperforming the next-best system by 29% in the Llama3.1-8B interactive benchmark,true,"entry 6.0-0047 against entry 6.0-0004 (the second entry cited in the post, 8 x RTX PRO 6000), Llama 3.1-8B Interactive",+29%,"44,087.3 against 34,241.0 (+28.8%). The next-best entry with the same accelerator and count is 6.0-0004",CONSISTENT,,false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H5,HPE,HPE v6.0,surpassing Dell’s PowerEdge XE7745 in six separate Llama scenarios,true,"Llama rows shared by entries 6.0-0047 and 6.0-0023 (Dell XE7745, 8 x RTX PRO 6000)",six Llama scenarios,"6 of 6 shared Llama rows are higher for HPE: llama2-70b-99 Offline 29,908 against 26,830, llama2-70b-99 Server 28,728 against 25,536, llama2-70b-99.9 Offline 29,900 against 26,830, llama2-70b-99.9 Server 28,717 against 25,536, llama3.1-8b Offline 49,580 against 49,036, llama3.1-8b Server 48,794 against 45,009",CONSISTENT,,false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H8,HPE,HPE v6.0,"including five No. 1 results and two solo results, when compared to other systems featuring 8 NVIDIA Blackwell / B200 SXM GPUs",true,"entry 6.0-0048 (XD685, 8 x B200 SXM) ranked among v6.0 rows with the same model, scenario, accelerator and count",five No. 1 and two solo,First among at least two entries: 5 of 5 listed rows. Only entry in its group: 2 of 2 listed rows,CONSISTENT,,false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H9,HPE,HPE v6.0,"In five scenarios, the HPE ProLiant Compute XD685 surpassed Supermicro SYS-822GS-NBRT results",true,entry 6.0-0048 against entry 6.0-0099 (Supermicro SYS-822GS-NBRT),"five scenarios, Whisper Offline 46033.9 against 42389.7","5 of 5 rows higher for HPE. Whisper Offline 46,033.9 against 42,389.7",CONSISTENT,,false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H1,HPE,HPE v6.0,claimed 18 No. 1 rankings across a wide spectrum of AI workloads,true,"HPE closed datacenter rows in v6.0 (78), ranked within rows of the same model, scenario, accelerator model and accelerator count, and ranked across all rows of the same model and scenario",18 No. 1 rankings,"First in the same-accelerator, same-count group: 55 rows. The same count when the group has at least two entries: 20. First across every system for its model and scenario: 2",CANNOT_TELL,The post does not say how a ranking is defined. No simple rule gives 18.,false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H12,HPE,HPE v6.0,"the power of the GPUs in the DL385 Gen11 was limited to a maximum of 450 watts, and the comparable Cisco UCS X215c M8 utilized GPUs operating at higher power (600 watts). The DL385 Gen11 produced almost identical inference results across key benchmarks, demonstrating superior performance-per-watt economics",true,power fields of entries 6.0-0054 and 6.0-0014,"450 W against 600 W, superior performance per watt",Neither entry has power results (has_power is false for both) and the summary file has no GPU power limit,CANNOT_TELL,Performance per watt needs a power measurement. The file has none for these two entries. MLCommons states that power figures such as a TDP are not measured or validated by MLCommons.,false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H2,HPE,HPE v6.0,submitted the highest total number of results—an impressive 862,true,rows per submitter in the v6.0 file (520 rows),"highest total, 862","Rows per submitter: HPE 86, Dell 68, NVIDIA 63. HPE has 14 entries",CANNOT_TELL,HPE has the most rows of any submitter. The file lists 86 rows for HPE and we cannot reproduce 862. The post does not say what it counts.,false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H6,HPE,HPE v6.0,"the HPE ProLiant DL385 Gen11, equipped with four NVIDIA RTX PRO™ 6000 Blackwell Server Edition GPUs, secured four No. 1 spots in Llama benchmarks",true,"the four benchmarks listed in footnote 6 (Llama 3.1-8B Offline and Interactive, Llama 2 70B 99.9 Server and Offline), for the four-GPU entry 6.0-0054 and for the two-GPU entry 6.0-0052 that the footnote cites","four No. 1 spots, four GPUs",First in group for the four-GPU entry 6.0-0054: 3 of 4. First in group for the two-GPU entry 6.0-0052: 4 of 4. Two of the four rows are sole entrants in their group,CANNOT_TELL,"The footnote cites 6.0-0052, which has two GPUs. The text says four GPUs. The result depends on which entry is meant.",false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+H7,HPE,HPE v6.0,delivered 24% higher performance than the next-best submission,true,Llama 3.1-8B Interactive per GPU: entries 6.0-0052 and 6.0-0054 against 6.0-0004 (8 GPUs),+24%,"6.0-0052 (2 GPUs) 5,295 per GPU, +23.7% against 6.0-0004 at 4,280. 6.0-0054 (4 GPUs) 5,030 per GPU, +17.5%",CANNOT_TELL,"The 24% reproduces for the two-GPU entry that the footnote cites and not for the four-GPU entry the text names. Per GPU, entry 6.0-0047 is higher than both.",false,HPE Community blog: HPE secures 18 top rankings in MLPerf Inference v6.0 AI benchmarks (post dated 1 April 2026),https://community.hpe.com/t5/ai-unlocked/hpe-secures-18-top-rankings-in-mlperf-inference-v6-0-ai/ba-p/7264016,2026-10-28,MLPerf Inference v6.0
+T8,AMD,AMD Training v6.0,"Across the partner submissions, results landed within 6% of official AMD submissions across both major LLM training workloads",true,"eight partner 8 x MI355X entries (Oracle, Dell x2, HPE, Supermicro, MiTAC, Vultr, GigaComputing) against the AMD 8 x MI355X entry, time to train, both benchmarks",within 6%,"Largest difference 6.9% (MiTAC, llama31_8b, 92.829 min against 86.845 min). Llama 2 70B largest 6.6%. Llama 3.1 8B largest 6.9%",BELOW,Every partner entry is slower than the AMD entry. MI350X partner entries and non-listed submitters are not in this set.,false,"AMD Blogs: AMD Instinct GPUs Deliver Breakthrough Performance, Multi-node Scale and Ecosystem Validation in MLPerf Training 6.0 (post dated 16 June 2026)",https://www.amd.com/en/blogs/2026/amd-delivers-breakthrough-mlperf-training-6-0-results.html,2026-10-28,MLPerf Training v6.0
+T1,AMD,AMD Training v6.0,"In just over one year, AMD improved performance by 3.5X from its first AMD Instinct MI300X GPU submission using MXFP8 in MLPerf Training 5.0 to the AMD Instinct MI355X GPU submission using MXFP4 in MLPerf Training 6.0.",true,"AMD 8-GPU Llama 2 70B LoRA time to train, round 5.0 MI300X entry against round 6.0 MI355X entry (official-style time: olympic average times the committee scaling factor)",3.5X,29.250 min against 8.271 min = 3.54x,CONSISTENT,Precision recipes are not in the result files.,false,"AMD Blogs: AMD Instinct GPUs Deliver Breakthrough Performance, Multi-node Scale and Ecosystem Validation in MLPerf Training 6.0 (post dated 16 June 2026)",https://www.amd.com/en/blogs/2026/amd-delivers-breakthrough-mlperf-training-6-0-results.html,2026-10-28,MLPerf Training v6.0
+T2,AMD,AMD Training v6.0,"Llama 2-70B fine-tuning: MI355X improved by 19%, while MI350X improved by 16%.",true,"AMD 8-GPU Llama 2 70B LoRA, round 5.1 entries against round 6.0 entries, reduction in time to train",19% and 16%,"MI355X 18.8% shorter, MI350X 16.2% shorter",CONSISTENT,The percentages reproduce as reductions in time to train. As speedups they would be 23% and 19%.,false,"AMD Blogs: AMD Instinct GPUs Deliver Breakthrough Performance, Multi-node Scale and Ecosystem Validation in MLPerf Training 6.0 (post dated 16 June 2026)",https://www.amd.com/en/blogs/2026/amd-delivers-breakthrough-mlperf-training-6-0-results.html,2026-10-28,MLPerf Training v6.0
+T3,AMD,AMD Training v6.0,"Llama 3.1-8B pre-training: MI355X improved by 13%, while MI350X improved by 11%.",true,"AMD 8-GPU Llama 3.1 8B, round 5.1 entries against round 6.0 entries, reduction in time to train",13% and 11%,"MI355X 12.9% shorter, MI350X 10.7% shorter",CONSISTENT,Reductions in time to train.,false,"AMD Blogs: AMD Instinct GPUs Deliver Breakthrough Performance, Multi-node Scale and Ecosystem Validation in MLPerf Training 6.0 (post dated 16 June 2026)",https://www.amd.com/en/blogs/2026/amd-delivers-breakthrough-mlperf-training-6-0-results.html,2026-10-28,MLPerf Training v6.0
+T4,AMD,AMD Training v6.0,coming within 5% on Llama 2-70B fine-tuning and within 6% on Llama 3.1-8B pre-training,true,"AMD 8 x MI355X round 6.0 entries against the fastest 8 x B200 SXM entries in round 6.0 (not GB200), throughput as the inverse of time to train",within 5% and within 6%,"Llama 2 70B: fastest B200 entry Supermicro 7.856 min, AMD 8.271 min, AMD throughput 95.0% of it. Llama 3.1 8B: fastest B200 entry Dell 82.213 min, AMD 86.845 min, 94.7%",CONSISTENT,The statement does not name the B200 entry. This uses the fastest in the file. Against the slower B200 entries AMD is ahead. Precision (MXFP4 and NVFP4) is not in the file.,false,"AMD Blogs: AMD Instinct GPUs Deliver Breakthrough Performance, Multi-node Scale and Ecosystem Validation in MLPerf Training 6.0 (post dated 16 June 2026)",https://www.amd.com/en/blogs/2026/amd-delivers-breakthrough-mlperf-training-6-0-results.html,2026-10-28,MLPerf Training v6.0
+T5,AMD,AMD Training v6.0,"AMD submitted FLUX.1 at 64 Instinct MI325X GPUs, while Oracle Cloud Infrastructure, in collaboration with AMD, submitted FLUX.1 at 512 GPUs.",true,"FLUX.1 entries from AMD and Oracle in round 6.0, GPU counts and accelerators",64 MI325X GPUs and 512 GPUs,"AMD 64 x AMD Instinct MI325X 256GB HBM3e, Oracle 512 x AMD Instinct MI300X 192GB HBM3",CONSISTENT,,false,"AMD Blogs: AMD Instinct GPUs Deliver Breakthrough Performance, Multi-node Scale and Ecosystem Validation in MLPerf Training 6.0 (post dated 16 June 2026)",https://www.amd.com/en/blogs/2026/amd-delivers-breakthrough-mlperf-training-6-0-results.html,2026-10-28,MLPerf Training v6.0
+T6,AMD,AMD Training v6.0,"With the 512-GPU OCI result, AMD matched NVIDIA's largest scale with a 512-GPU Training submission.",true,"largest NVIDIA GPU count in round 6.0, for FLUX.1 and for any benchmark",matched NVIDIA at 512 GPUs,NVIDIA largest FLUX.1 entry: 512 GPUs. NVIDIA largest entry in any benchmark: 8192 GPUs,CONSISTENT,It matches for FLUX.1. NVIDIA entries in other benchmarks are larger.,false,"AMD Blogs: AMD Instinct GPUs Deliver Breakthrough Performance, Multi-node Scale and Ecosystem Validation in MLPerf Training 6.0 (post dated 16 June 2026)",https://www.amd.com/en/blogs/2026/amd-delivers-breakthrough-mlperf-training-6-0-results.html,2026-10-28,MLPerf Training v6.0
+T7,AMD,AMD Training v6.0,This round included a record 10 ecosystem partners submitting MLPerf Training results on AMD Instinct GPU-based platforms,true,"submitters other than AMD with an AMD Instinct accelerator in rounds 5.0, 5.1 and 6.0",a record 10 partners,"Round 5.0: 7. Round 5.1: 9. Round 6.0: 10 (Cisco, Dell, GigaComputing, HPE, Krai, MiTAC, Oracle, Supermicro, tinycorp, vultr)",CONSISTENT,The list in the post names MiTAC and MiTAC + Akash separately and omits tinycorp. The file has one MiTAC submitter and a tinycorp entry. Rounds before 5.0 are not read.,false,"AMD Blogs: AMD Instinct GPUs Deliver Breakthrough Performance, Multi-node Scale and Ecosystem Validation in MLPerf Training 6.0 (post dated 16 June 2026)",https://www.amd.com/en/blogs/2026/amd-delivers-breakthrough-mlperf-training-6-0-results.html,2026-10-28,MLPerf Training v6.0
+LM1,Lambda,Lambda Training v6.0,"Our GB300 NVL72 system converges in 11.59 minutes, compared to 14.25 minutes in Training v5.1. This represents an 18.7% improvement in training speed",true,"Lambda 72-GPU GB300 Llama 3.1 8B, round 5.1 entry against round 6.0 entry, official-style time to train","11.59 and 14.25 minutes, 18.7%","11.586 min against 14.255 min, 18.7% shorter (1.23x)",CONSISTENT,The 18.7% is the reduction in time to train. As a speedup it is 23.0%.,false,Lambda: MLPerf Training v6.0: Lambda delivers fastest LLM training on NVIDIA GB300 NVL72 and fastest MoE training on NVIDIA HGX B200 (post dated 16 June 2026),https://lambda.ai/blog/mlperf-training-v6.0-lambda-delivers-fastest-llm-training,2026-10-28,MLPerf Training v6.0
+LM2,Lambda,Lambda Training v6.0,"The results table in the post lists Lambda Llama 3.1 8B at 11.59 minutes on GB300 NVL72 and 85.26 minutes on HGX B200, and GPT-OSS-20B at 18.35 minutes on GB300 NVL72 and 96.46 minutes on HGX B200.",false,Lambda entries in round 6.0,"11.59, 85.26, 18.35 and 96.46 minutes",llama31_8b gb300 11.586. llama31_8b b200 85.263. gpt_oss_20b gb300 18.359. gpt_oss_20b b200 96.457,CONSISTENT,The GPT-OSS-20B GB300 time is 18.359 in the log summary against 18.35 in the post.,false,Lambda: MLPerf Training v6.0: Lambda delivers fastest LLM training on NVIDIA GB300 NVL72 and fastest MoE training on NVIDIA HGX B200 (post dated 16 June 2026),https://lambda.ai/blog/mlperf-training-v6.0-lambda-delivers-fastest-llm-training,2026-10-28,MLPerf Training v6.0
+LM3,Lambda,Lambda Training v6.0,"Lambda completed training in 11.59 minutes on a GB300 NVL72 system, the fastest result among all GB300 NVL72 systems submitted this round, edging the NVIDIA Theia system single-rack submission (11.75 min), HPE (11.82 min), GigaComputing (11.84 min), and Nebius (11.87 min).",true,every 72-GPU GB300 Llama 3.1 8B entry in round 6.0,"fastest, with 11.75, 11.82, 11.84 and 11.87 behind","Lambda 11.586, NVIDIA 11.753, HPE 11.821, GigaComputing 11.844, Supermicro 11.860, Nebius 11.872",CONSISTENT,NVIDIA also has a 512-GPU GB300 entry at 4.636 min. The comparison is among single-rack entries. A Supermicro entry at 11.860 also sits behind Lambda and is not named in the post.,false,Lambda: MLPerf Training v6.0: Lambda delivers fastest LLM training on NVIDIA GB300 NVL72 and fastest MoE training on NVIDIA HGX B200 (post dated 16 June 2026),https://lambda.ai/blog/mlperf-training-v6.0-lambda-delivers-fastest-llm-training,2026-10-28,MLPerf Training v6.0
+LM4,Lambda,Lambda Training v6.0,Lambda achieved the fastest result among single-node HGX B200 submissions for GPT-OSS-20B.,true,8-GPU B200 GPT-OSS-20B entries in round 6.0,fastest among single-node HGX B200,"Lambda 96.457, Netweb_Technologies_India_LTD 97.835, Dell 98.595, Supermicro 98.773",CONSISTENT,The margin over the next entry is 1.4%.,false,Lambda: MLPerf Training v6.0: Lambda delivers fastest LLM training on NVIDIA GB300 NVL72 and fastest MoE training on NVIDIA HGX B200 (post dated 16 June 2026),https://lambda.ai/blog/mlperf-training-v6.0-lambda-delivers-fastest-llm-training,2026-10-28,MLPerf Training v6.0
+LM7,Lambda,Lambda Training v6.0,Llama 3.1 8B was introduced in the previous round (MLPerf Training v5.1),true,"benchmarks present in rounds 5.0, 5.1 and 6.0",introduced in v5.1,Llama 3.1 8B in round 5.0: no. In round 5.1: yes,CONSISTENT,,false,Lambda: MLPerf Training v6.0: Lambda delivers fastest LLM training on NVIDIA GB300 NVL72 and fastest MoE training on NVIDIA HGX B200 (post dated 16 June 2026),https://lambda.ai/blog/mlperf-training-v6.0-lambda-delivers-fastest-llm-training,2026-10-28,MLPerf Training v6.0
+LM8,Lambda,Lambda Training v6.0,"GPT-OSS-20B, along with DeepSeek-V3, is one of MLPerf's first mixture-of-experts benchmarks.",true,"benchmarks present in rounds 5.0, 5.1 and 6.0",first mixture-of-experts benchmarks,Both appear in round 6.0 only. Present in round 5.0: no. Present in round 5.1: no. Present in round 6.0: yes,CONSISTENT,Rounds before 5.0 are not read.,false,Lambda: MLPerf Training v6.0: Lambda delivers fastest LLM training on NVIDIA GB300 NVL72 and fastest MoE training on NVIDIA HGX B200 (post dated 16 June 2026),https://lambda.ai/blog/mlperf-training-v6.0-lambda-delivers-fastest-llm-training,2026-10-28,MLPerf Training v6.0
+LM5,Lambda,Lambda Training v6.0,"Lambda converged in 18.35 minutes, among the fastest of the GB300 NVL72 single-rack systems.",true,every 72-GPU GB300 GPT-OSS-20B entry in round 6.0,among the fastest,"Rank 6 of 6: GigaComputing 18.055, HPE 18.126, Supermicro 18.186, Nebius 18.196, NVIDIA 18.290, Lambda 18.359. Lambda is 1.7% behind the first",CANNOT_TELL,The statement is a judgment. The spread across the group is under 2%.,false,Lambda: MLPerf Training v6.0: Lambda delivers fastest LLM training on NVIDIA GB300 NVL72 and fastest MoE training on NVIDIA HGX B200 (post dated 16 June 2026),https://lambda.ai/blog/mlperf-training-v6.0-lambda-delivers-fastest-llm-training,2026-10-28,MLPerf Training v6.0
+LM6,Lambda,Lambda Training v6.0,"Eight teams reported official GB300 NVL72 results this round: Lambda, Nebius, CoreWeave, Oracle, Dell, GigaComputing, HPE, and NVIDIA.",true,submitters with a GB300 accelerator in round 6.0,eight teams,"10 submitters: CoreWeave, Dell, GigaComputing, HPE, Lambda, NVIDIA, Nebius, Oracle, Quanta_Cloud_Technology, Supermicro",CANNOT_TELL,Supermicro and Quanta Cloud Technology are in the file with GB300 entries and are not on the list. Whether their systems count as NVL72 is not stated in the file.,false,Lambda: MLPerf Training v6.0: Lambda delivers fastest LLM training on NVIDIA GB300 NVL72 and fastest MoE training on NVIDIA HGX B200 (post dated 16 June 2026),https://lambda.ai/blog/mlperf-training-v6.0-lambda-delivers-fastest-llm-training,2026-10-28,MLPerf Training v6.0
+CW1,CoreWeave,CoreWeave Training v6.0,"CoreWeave trained DeepSeek-V3 671B to the benchmark's quality target in 2.02 minutes on 8,192 NVIDIA GB300 NVL72 GPUs. It was the fastest result in the round on that benchmark.",true,CoreWeave DeepSeek-V3 671B entries and every other DeepSeek-V3 entry in round 6.0,"2.02 minutes on 8,192 GPUs, fastest in the round","2.021 min on 8192 GPUs. Fastest in the round: CoreWeave 2.021 min on 8192 GPUs. Next: CoreWeave 3.092 (4096 GPUs), NVIDIA 3.340 (8192 GPUs)",CONSISTENT,,false,"CoreWeave: Scaling frontier-class AI training on CoreWeave, a deep technical analysis of MLPerf Training v6.0 results (whitepaper page)",https://coreweave.com/resources/whitepapers/scaling-frontier-class-ai-training-on-coreweave-a-deep-technical-analysis-of-mlperf-r-training-v6-0-results,2026-10-28,MLPerf Training v6.0
+CW2,CoreWeave,CoreWeave Training v6.0,"A year-over-year read on Llama 3.1 405B, where time-to-train dropped 2.8x, from 27.33 minutes to 9.77 minutes",true,round 5.0 entry at 27.33 minutes and CoreWeave round 6.0 Llama 3.1 405B entry,"2.8x, 27.33 to 9.77 minutes",IBM+CoreWeave+NVIDIA carina_ngpu2496_ngc25.04_nemo (2496 GPUs) 27.335 min against CoreWeave 4096 GPUs 9.771 min = 2.80x,CONSISTENT,"The earlier entry is a joint submission (IBM, CoreWeave, NVIDIA) on 2,496 GPUs. The later one uses 4,096 GPUs.",false,"CoreWeave: Scaling frontier-class AI training on CoreWeave, a deep technical analysis of MLPerf Training v6.0 results (whitepaper page)",https://coreweave.com/resources/whitepapers/scaling-frontier-class-ai-training-on-coreweave-a-deep-technical-analysis-of-mlperf-r-training-v6-0-results,2026-10-28,MLPerf Training v6.0
+CW3,CoreWeave,CoreWeave Training v6.0,"It walks through how CoreWeave scaled frontier training across five system configurations, from 64 GPUs to 8,192",true,CoreWeave entries in round 6.0,"five system configurations, 64 to 8,192 GPUs","5 systems, GPU counts [64, 2048, 4096, 8192]",CONSISTENT,,false,"CoreWeave: Scaling frontier-class AI training on CoreWeave, a deep technical analysis of MLPerf Training v6.0 results (whitepaper page)",https://coreweave.com/resources/whitepapers/scaling-frontier-class-ai-training-on-coreweave-a-deep-technical-analysis-of-mlperf-r-training-v6-0-results,2026-10-28,MLPerf Training v6.0
+B4,NVIDIA,NVIDIA v6.1,reaching 0.65 720p videos per second at 5.7 seconds per video,true,"NVIDIA GB300 NVL72 (72 GPUs) WAN 2.2, Offline and SingleStream",0.65 videos/s and 5.7 s per video,Offline 0.6546 samples/s. SingleStream 5.808 s on GB300. The NVIDIA GB200 NVL72 row is 5.732 s,BELOW,Throughput matches. The GB300 latency is 5.8 s. The 5.7 s figure matches the GB200 row.,false,NVIDIA Blog: NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut,https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference,2026-10-28,MLPerf Inference v6.1
+B1,NVIDIA,NVIDIA v6.1,"Vera Rubin NVL72 delivers up to 3.7x higher throughput than GB300 NVL72 on Qwen3-VL across offline, server and interactive scenarios",true,"NVIDIA Vera Rubin NVL72 entry 6.1-0106 (72 GPUs, preview) against GB300 NVL72 entry 6.1-0074 (72 GPUs), Qwen3-VL-235B",up to 3.7x,"Offline 1.83x, Server 1.92x, Interactive 3.74x",CONSISTENT,The maximum is the Interactive scenario. The Vera Rubin entry has availability preview.,false,NVIDIA Blog: NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut,https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference,2026-10-28,MLPerf Inference v6.1
+B2,NVIDIA,NVIDIA v6.1,"On DeepSeek-R1, using the NVIDIA TensorRT-LLM library, throughput is up to 2.5x higher than GB300 NVL72.",true,"same two entries, DeepSeek-R1",up to 2.5x,"Offline 1.74x, Server 1.97x, Interactive 2.57x",CONSISTENT,The maximum is the Interactive scenario. The statement rounds down.,false,NVIDIA Blog: NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut,https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference,2026-10-28,MLPerf Inference v6.1
+B3,NVIDIA,NVIDIA v6.1,"NVIDIA's DeepSeek-R1 (DSR1) submission scaled from a single GB300 NVL72 rack (72 GPUs) to four racks (288 GPUs), achieving 99% scaling efficiency in the offline scenario.",true,"NVIDIA GB300 NVL72 entries 6.1-0073 (288 GPUs) and 6.1-0074 (72 GPUs), DeepSeek-R1 Offline",99%,"2,705,130 against 4 x 679,740 = 99.5%",CONSISTENT,,false,NVIDIA Blog: NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut,https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference,2026-10-28,MLPerf Inference v6.1
+B6,NVIDIA,NVIDIA v6.1,"In v6.1, GB300 NVL72 performance on Qwen3-VL improved up to 1.6x over v6.0 results.",true,NVIDIA GB300 Qwen3-VL-235B Server per GPU: v6.1 entry 6.1-0074 (72 GPUs) against v6.0 entry 6.0-0077 (4 GPUs),up to 1.6x,16.81 per GPU against 10.41 (1.62x),CONSISTENT,"Per GPU, Server scenario. The v6.0 entry has 4 GPUs. Offline per GPU is lower.",false,NVIDIA Blog: NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut,https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference,2026-10-28,MLPerf Inference v6.1
+B5,NVIDIA,NVIDIA v6.1,9x higher throughput and 7.5x lower latency than a single node,true,NVIDIA GB300 NVL72 (72 GPUs) WAN 2.2 against the smaller entries in the file: Oracle GB300 4-GPU and NVIDIA DGX B300 8-GPU,"9x throughput, 7.5x latency","Against the 4-GPU GB300 entry: 47.3x throughput, 5.3x lower latency. Against the 8-GPU DGX B300 entry: 10.5x and 3.1x",CANNOT_TELL,Neither smaller entry gives 9x and 7.5x. The single-node baseline the statement uses is not in the file.,false,NVIDIA Blog: NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut,https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference,2026-10-28,MLPerf Inference v6.1
+B7,NVIDIA,NVIDIA v6.1,"The NVIDIA partner ecosystem participated broadly, with 19 partners ... eight of them on multi-node Blackwell NVL72 systems",true,"organizations other than NVIDIA with an NVIDIA accelerator in the v6.1 file, and those with a multi-node GB200 or GB300 system","19 partners, eight on multi-node NVL72","All 19 named organizations have NVIDIA entries. The file has 26 such organizations in all. Multi-node GB200 or GB300 systems: 8 organizations (Azure, CoreWeave, Crusoe, Dell, Google, HPE, Nebius, Oracle). Among the 19 named: 7",CANNOT_TELL,"Eight organizations in the file have multi-node systems, but one of them, Google, is not on the post list. Among the 19 named the count is seven. The statement does not say what counts as multi-node or as a partner.",false,NVIDIA Blog: NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut,https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference,2026-10-28,MLPerf Inference v6.1
+B8,NVIDIA,NVIDIA v6.1,Vera Rubin NVL72 delivered 30 x better performance than GB300 NVL72 in preview testing.,true,SemiAnalysis AgentX results are not MLPerf results,30x on AgentX,The v6.1 file holds no AgentX benchmark,CANNOT_TELL,The statement refers to a different benchmark that is not in the MLCommons files.,false,NVIDIA Blog: NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut,https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference,2026-10-28,MLPerf Inference v6.1
+AV10,AMD,AMD v6.1,"The system delivered 285,454 Offline tokens per second and 253,501 Server tokens per second.",true,Dell MangoBoost four-cluster GPT-OSS-120B entry,"285,454 and 253,501","285,454 and 253,501",CONSISTENT,The regions (Korea and the United States) are not in the file.,false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV11,AMD,AMD v6.1,"The average of comparable AMD MI355X GPU partner results landed within 4% of AMD, and some matched or slightly exceeded the AMD reference result.",true,"partner 8-GPU MI355X entries (Dell, Dell MangoBoost, HPE, MiTAC, Oracle, Supermicro) against the AMD 8-GPU entry, same model and scenario, Offline, Server and Interactive","average within 4%, some at or above AMD",31 comparisons. Mean 97.4% of the AMD value (range 94.6% to 101.6%). 4 at or above 100%,CONSISTENT,Dell_AMD MI350P and MI350X entries are not in this set.,false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV12,AMD,AMD v6.1,"Dell Technologies, Oracle, Hewlett Packard Enterprise, Supermicro, MangoBoost, Crusoe, and MiTAC submitted results across",true,organizations other than AMD with an AMD Instinct accelerator in the v6.1 file,seven partners,"10 organizations: Cisco, Crusoe, Dell, Dell_AMD, Dell_MangoBoost, HPE, MangoBoost, MiTAC, Oracle, Supermicro",CONSISTENT,All seven named organizations are in the file. Cisco and the Dell_AMD and Dell_MangoBoost labels add to the count.,false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV13,AMD,AMD v6.1,the AMD Instinct MI350P PCIe ... card submitted five Closed workloads,true,Dell_AMD MI350P entries in the v6.1 file,five Closed workloads,"5 models: dlrm-v3, gpt-oss-120b, llama2-70b-99.9, llama3.1-8b, wan-2.2-t2v-a14b",CONSISTENT,,false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV14,AMD,AMD v6.1,GPU platform retained approximately 80% of the AMD Instinct ... MI355X platform performance,true,"AMD 8-GPU MI350X entries against the AMD 8-GPU MI355X entry: DLRM v3 Offline and Server, Wan-2.2 Offline and SingleStream (latency inverted)",about 80%,"83.6%, 76.2%, 77.3%, 78.3%. Mean 78.8%",CONSISTENT,"The statement says selected workloads. GPT-OSS, Llama and rated TDP are not compared here. The MI350X entries come from different AMD systems.",false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV2,AMD,AMD v6.1,"increased GPT-OSS-120B throughput by 28% in Offline and 38% in Server, while Wan-2.2 SingleStream performance increased 70%",true,"AMD 8-GPU MI355X entry, v6.0 entries 6.0-0002 and 6.0-0103 against the v6.1 entry","+28% Offline, +38% Server, Wan-2.2 SingleStream +70%","GPT-OSS-120B Offline 95,004 to 121,818 (+28.2%), Server 82,136 to 113,241 (+37.9%). Wan-2.2 SingleStream 27.38 s to 16.12 s, +69.8% faster, 41.1% shorter",CONSISTENT,The 70% is a speedup (the latency is 41% shorter). The v6.0 Wan-2.2 entry is in the open division. The v6.1 entry is closed.,false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV3,AMD,AMD v6.1,72 AMD Instinct MI355X GPUs in MLPerf Inference round 6.1 delivered more GPT-OSS-120B throughput than 94 GPUs deliveredin previous round 6.0.,true,"AMD 72-GPU v6.1 entry against AMD v6.0 entry 6.0-0003 (94 GPUs by its accelerator name, 96 in the total column)",72 GPUs in 6.1 above 94 GPUs in 6.0,"Offline 1,039,900 against 1,031,070 (+0.9%). Server 964,468 against 900,054 (+7.2%)",CONSISTENT,The v6.0 entry total column says 96 accelerators. Its accelerator name says 94.,false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV4,AMD,AMD v6.1,"AMD expanded GPT-OSS-120B from 8 to 72 AMD Instinct MI355X GPUs, a ninefold increase in GPU count, while retaining 95% scale efficiency in both Offline and Server.",true,"AMD 72-GPU entry against 9 times the AMD 8-GPU entry, GPT-OSS-120B",95% in both,"Offline 1,039,900 against 9 x 121,818 = 94.9%. Server 964,468 against 9 x 113,241 = 94.6%",CONSISTENT,,false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV5,AMD,AMD v6.1,"Through Crusoe, GPT-OSS-120B reached 5.75 million tokens per second in Offline and 5.39 million tokens per second in Server, the highest aggregate token throughput ever submitted in MLPerf history.",true,"Crusoe 512-GPU GPT-OSS-120B entry against the largest tokens-per-second value in rounds 5.0, 5.1, 6.0 and 6.1","5.75 million and 5.39 million, the highest ever","Offline 5,749,440, Server 5,392,930. Largest in v6.1: Crusoe 5,749,440. Largest earlier: Supermicro 129,047. AMD_MangoBoost 648,248. NVIDIA 2,494,310",CONSISTENT,Rounds before 5.0 are not read.,false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV6,AMD,AMD v6.1,"The DeepSeek-R1 submission reached 2.90 million Offline and 2.41 million Server tokens per second, establishing the highest aggregate token throughput on DeepSeek-R1 in MLPerf history.",true,"Crusoe 512-GPU DeepSeek-R1 entry against every other DeepSeek-R1 entry in rounds 5.1, 6.0 and 6.1","2.90 million and 2.41 million, the highest on DeepSeek-R1","Offline 2,901,950, Server 2,405,310. Largest other DeepSeek-R1 result: 2,705,130 (v6.1), 2,494,310 (earlier rounds)",CONSISTENT,,false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV7,AMD,AMD v6.1,The AMD Instinct MI355X GPU then delivered 18% higher Offline and 11% higher SingleStream performance than the selected NVIDIA B300 submission.,true,AMD 8-GPU MI355X Wan-2.2 entry against the NVIDIA DGX B300 (8 GPUs) entry,"+18% Offline, +11% SingleStream",Offline 0.0731 against 0.0622 (+17.6%). SingleStream 16.124 s against 17.977 s (+11.5% faster). VibeHPC 8-GPU B300 Offline is 0.0784,CONSISTENT,"The statement says selected. Another B300 entry, from VibeHPC, is above the AMD Offline value.",false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV8,AMD,AMD v6.1,"On a standard AMD Instinct MI355X 8-GPU system, AMD led the selected NVIDIA B200 and B300 results in both Offline and Server.",true,AMD 8-GPU GPT-OSS-120B entry against every 8-GPU B200 and B300 entry in the v6.1 file,AMD ahead of the selected B200 and B300 entries,"Offline: AMD 121,818, best B200 or B300 entry HPE 114,569. Server: AMD 113,241, best entry HPE 113,685",CONSISTENT,AMD is ahead of every B200 and B300 8-GPU entry in Offline. In Server one B300 entry from HPE is 0.4% higher. The statement says selected.,false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV9,AMD,AMD v6.1,AMD Instinct MI355X GPUs led NVIDIA B200 and B300 GPT-OSS-120B results at 8 GPUs and the NVIDIA GB200 result at 72 GPUs.,true,AMD 72-GPU entry against the 72-GPU GB200 GPT-OSS-120B entries in the v6.1 file,AMD ahead of the GB200 result at 72 GPUs,"Offline 1,039,900 against best GB200 911,566 (+14.1%). Server 964,468 against 901,058 (+7.0%)",CONSISTENT,The GB300 72-GPU entry is higher than the AMD value.,false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+AV1,AMD,AMD v6.1,"expanding from three model families in the prior MLPerf 6.0 Inference round, to now six, across AMD Instinct",true,"closed models on AMD accelerators in round 6.1, AMD own entries and all entries",three families to six,"AMD own v6.1 entries: 4 models (dlrm-v3, gpt-oss-120b, llama2-70b-99.9, wan-2.2-t2v-a14b). All v6.1 entries on AMD accelerators: 6 models (deepseek-r1, dlrm-v3, gpt-oss-120b, llama2-70b-99.9, llama3.1-8b, wan-2.2-t2v-a14b)",CANNOT_TELL,"The file does not define a model family. Counts are four, six or seven depending on whose entries and which models are counted.",false,AMD Blogs: AMD Delivers Its Broadest MLPerf Inference 6.1 Submission (post dated 16 September 2026),https://www.amd.com/en/blogs/2026/amd-delivers-its-broadest-mlperf-inference-6-1-submission.html,2026-10-28,MLPerf Inference v6.1
+CB1,Crusoe,Crusoe v6.1,"Crusoe's MLPerf Inference v6.1 submission ran gpt-oss-120b and DeepSeek-R1 on 512 AMD Instinct MI355X GPUs, the largest MI355X inference entry by GPU count in MLPerf history.",true,MI355X entries in rounds 6.0 and 6.1 by GPU count,"512 GPUs, the largest MI355X entry",Crusoe 512 GPUs. Largest other MI355X entry in v6.1: 72 GPUs,CONSISTENT,,false,Crusoe: Serving 5.75 million tokens per second: Crusoe's MLPerf Inference v6.1 results on AMD MI355X (post dated 16 September 2026),https://www.crusoe.ai/resources/blog/serving-5-75-million-tokens-per-second-crusoes-mlperf-inference-v6-1-results-on-amd-mi355x,2026-10-28,MLPerf Inference v6.1
+CB2,Crusoe,Crusoe v6.1,"The results table in the post lists the Crusoe 512-GPU entries at about 5.75M, 5.39M, 2.90M and 2.40M tokens per second in total and about 11.2k, 10.5k, 5.7k and 4.7k per GPU, for GPT-OSS-120B and DeepSeek-R1, Offline and Server.",false,"Crusoe 512-GPU entries, total divided by 512","5.75M, 5.39M, 2.90M and 2.40M tokens/s, 11.2k, 10.5k, 5.7k and 4.7k per GPU","GPT-OSS-120B Offline 5,749,440 (11,229 per GPU), Server 5,392,930 (10,533). DeepSeek-R1 Offline 2,901,950 (5,668), Server 2,405,310 (4,698)",CONSISTENT,,false,Crusoe: Serving 5.75 million tokens per second: Crusoe's MLPerf Inference v6.1 results on AMD MI355X (post dated 16 September 2026),https://www.crusoe.ai/resources/blog/serving-5-75-million-tokens-per-second-crusoes-mlperf-inference-v6-1-results-on-amd-mi355x,2026-10-28,MLPerf Inference v6.1
+CB4,Crusoe,Crusoe v6.1,"Both models cleared closed-division accuracy: gpt-oss-120b at 83.7% exact-match against an 82.3% reference floor, and DeepSeek-R1 at 80.7% exact-match (80.54% floor) with a mean output length of 3,898 tokens per sample",true,accuracy fields of the Crusoe 512-GPU entries,"83.7, 80.7 and 3,898",Accuracy fields: exact_match: 80.69735642661804  TOKENS_PER_SAMPLE: 3898.373746581586. exact_match: 83.706,CONSISTENT,The reference floors and the output length range are not in the summary file.,false,Crusoe: Serving 5.75 million tokens per second: Crusoe's MLPerf Inference v6.1 results on AMD MI355X (post dated 16 September 2026),https://www.crusoe.ai/resources/blog/serving-5-75-million-tokens-per-second-crusoes-mlperf-inference-v6-1-results-on-amd-mi355x,2026-10-28,MLPerf Inference v6.1
+CB3,Crusoe,Crusoe v6.1,Throughput scales linearly from 8 to 512 GPUs at over 90% of ideal,true,"Crusoe 512-GPU GPT-OSS-120B entry against 64 times the AMD 8-GPU MI355X entry (the post baseline, a Crusoe 8-GPU run, is not in the file)",over 90% of ideal,"Offline 5,749,440 against 64 x 121,818 = 73.7%. Server 5,392,930 against 64 x 113,241 = 74.4%",CANNOT_TELL,"The Crusoe 8-GPU baseline is not in the file. Against the AMD 8-GPU entry the ratio is about 74%. Per GPU, the AMD entry is higher.",false,Crusoe: Serving 5.75 million tokens per second: Crusoe's MLPerf Inference v6.1 results on AMD MI355X (post dated 16 September 2026),https://www.crusoe.ai/resources/blog/serving-5-75-million-tokens-per-second-crusoes-mlperf-inference-v6-1-results-on-amd-mi355x,2026-10-28,MLPerf Inference v6.1
+CB5,Crusoe,Crusoe v6.1,(gpt-oss measured 2.69 s TTFT and 36.5 ms TPOT; DeepSeek 1.85 s TTFT and 79.98 ms TPOT),true,latency fields of the Crusoe entries,"TTFT 2.69 s and 1.85 s, TPOT 36.5 ms and 79.98 ms",The summary file has no time to first token or time per output token,CANNOT_TELL,"Latency percentiles are in the per-entry logs, not in the summary file.",false,Crusoe: Serving 5.75 million tokens per second: Crusoe's MLPerf Inference v6.1 results on AMD MI355X (post dated 16 September 2026),https://www.crusoe.ai/resources/blog/serving-5-75-million-tokens-per-second-crusoes-mlperf-inference-v6-1-results-on-amd-mi355x,2026-10-28,MLPerf Inference v6.1
+LV4,Lambda,Lambda v5.1,SDXL was 1.07x faster than the best 8xH200 results,true,Lambda 8 x B200 SDXL entry against the best 8 x H200 SDXL entry in v5.1,1.07x,"Offline 1.70x (Quanta_Cloud_Technology 19.19), Server 1.54x (Quanta_Cloud_Technology 18.42)",ABOVE,"The statement does not say which scenario. The 1.07x equals the Offline gain over the v5.0 best 8 x B200 entry (32.57 over 30.38), not the gain over the best 8 x H200 entry.",false,"Lambda: LLM performance up 15.4%: MLPerf v5.1 confirms NVIDIA HGX B200 on Lambda is built for enterprise inference (post dated 9 September 2025, updated 22 October 2026)",https://lambda.ai/blog/lambda-mlperf-inference-v5.1,2026-10-28,MLPerf Inference v5.1
+LV1,Lambda,Lambda v5.1,"The results table in the post lists Lambda v5.1 results against the v5.0 best for six rows: Llama 2 70B 99 Offline 102,725.00 against 98,858.00 and Server 99,993.90 against 98,443.30, Llama 3.1 405B Offline 1,648.60 against 1,538.17 and Server 1,246.79 against 1,080.31, Stable Diffusion XL Offline 32.57 against 30.38 and Server 28.46 against 28.92, with changes of +3.9%, +1.6%, +7.2%, +15.4%, +7.2% and -1.6%.",false,Lambda entry 5.1-0060 (8 x B200) against the best 8 x B200 entry in v5.0 for each row,"six rows, deltas +3.9, +1.6, +7.2, +15.4, +7.2 and -1.6 percent","llama2-70b-99 Offline: 102,725.00 against 98,858.00 (+3.9%). llama2-70b-99 Server: 99,993.90 against 98,443.30 (+1.6%). llama3.1-405b Offline: 1,648.60 against 1,538.17 (+7.2%). llama3.1-405b Server: 1,246.79 against 1,080.31 (+15.4%). stable-diffusion-xl Offline: 32.57 against 30.38 (+7.2%). stable-diffusion-xl Server: 28.46 against 28.92 (-1.6%)",CONSISTENT,"The v5.0 best is the best 8 x B200 entry. A MangoBoost 8-accelerator entry is higher in Llama 2 70B Offline at 103,182.",false,"Lambda: LLM performance up 15.4%: MLPerf v5.1 confirms NVIDIA HGX B200 on Lambda is built for enterprise inference (post dated 9 September 2025, updated 22 October 2026)",https://lambda.ai/blog/lambda-mlperf-inference-v5.1,2026-10-28,MLPerf Inference v5.1
+LV2,Lambda,Lambda v5.1,"Five of six submissions delivered higher performance than the prior rounds, with gains ranging from +1.6% to +15.4%",true,same six rows,"five of six higher, +1.6% to +15.4%",5 of 6 higher. Gains +1.6% to +15.4%,CONSISTENT,,false,"Lambda: LLM performance up 15.4%: MLPerf v5.1 confirms NVIDIA HGX B200 on Lambda is built for enterprise inference (post dated 9 September 2025, updated 22 October 2026)",https://lambda.ai/blog/lambda-mlperf-inference-v5.1,2026-10-28,MLPerf Inference v5.1
+LV3,Lambda,Lambda v5.1,"On 8xB200, both Llama 3.1 405B and Llama 2 70B delivered 3-4x higher performance than the best 8xH200 results",true,"Lambda 8 x B200 entry against the best 8 x H200 entry in v5.1, Llama 3.1 405B and Llama 2 70B 99",3-4x higher,"llama2-70b-99 Offline 2.91x (Dell), llama2-70b-99 Server 2.92x (ASUSTeK), llama3.1-405b Offline 2.98x (Nebius), llama3.1-405b Server 4.22x (Quanta_Cloud_Technology)",CONSISTENT,The ratios are 2.9 to 4.2. The two Llama 2 70B ratios are 2.9.,false,"Lambda: LLM performance up 15.4%: MLPerf v5.1 confirms NVIDIA HGX B200 on Lambda is built for enterprise inference (post dated 9 September 2025, updated 22 October 2026)",https://lambda.ai/blog/lambda-mlperf-inference-v5.1,2026-10-28,MLPerf Inference v5.1
+LV5,Lambda,Lambda v5.1,Host CPU : Dual Intel Xeon Platinum 8750 with 56 cores,true,host processor fields of Lambda entry 5.1-0060,"dual Xeon Platinum 8750, 56 cores","Intel(R) Xeon(R) Platinum 8570, 2 per node, 56 cores each",CONSISTENT,The post writes 8750. The file lists 8570.,false,"Lambda: LLM performance up 15.4%: MLPerf v5.1 confirms NVIDIA HGX B200 on Lambda is built for enterprise inference (post dated 9 September 2025, updated 22 October 2026)",https://lambda.ai/blog/lambda-mlperf-inference-v5.1,2026-10-28,MLPerf Inference v5.1
+LV7,Lambda,Lambda v5.1,using NVIDIA TensorRT 10.11 on NVIDIA CUDA 12.9,true,software field of Lambda entry 5.1-0060,"TensorRT 10.11, CUDA 12.9","TensorRT 10.11, CUDA 12.9",CONSISTENT,,false,"Lambda: LLM performance up 15.4%: MLPerf v5.1 confirms NVIDIA HGX B200 on Lambda is built for enterprise inference (post dated 9 September 2025, updated 22 October 2026)",https://lambda.ai/blog/lambda-mlperf-inference-v5.1,2026-10-28,MLPerf Inference v5.1
+LV6,Lambda,Lambda v5.1,"Lambda placed first or second in rankings across three MLPerf categories, standing out among a dozen top vendors",true,"Lambda ranks among 8 x B200 entries in v5.1, per model and scenario",first or second in three categories,llama2-70b-99 Offline: rank 2 of 10. llama2-70b-99 Server: rank 2 of 10. llama3.1-405b Offline: rank 6 of 13. llama3.1-405b Server: rank 3 of 13. stable-diffusion-xl Offline: rank 1 of 8. stable-diffusion-xl Server: rank 7 of 8,CANNOT_TELL,The post does not name its categories or its group of vendors. Lambda is first or second in three of the six rows within the 8 x B200 group.,false,"Lambda: LLM performance up 15.4%: MLPerf v5.1 confirms NVIDIA HGX B200 on Lambda is built for enterprise inference (post dated 9 September 2025, updated 22 October 2026)",https://lambda.ai/blog/lambda-mlperf-inference-v5.1,2026-10-28,MLPerf Inference v5.1
+NE1,Nebius,Nebius v5.1,"compared to the best results of the previous round, our one-host installation achieved 6.7% and 14.2% performance increases running Llama 3.1 405B inference in offline and server modes, respectively",true,"Nebius 4-GPU GB200 entry 5.1-0074 against the best 4-GPU GB200 entries in v5.0, Llama 3.1 405B","+6.7% Offline, +14.2% Server",Offline 855.82 against 801.91 (+6.7%). Server 596.11 against 522.12 (+14.2%),CONSISTENT,The v5.0 best is a 4-GPU GB200 entry from CoreWeave. A 72-GPU NVIDIA entry is far higher in both rounds.,false,Nebius: Nebius proves bare-metal-class performance for AI inference workloads in MLPerf Inference v5.1,https://nebius.com/blog/posts/bare-metal-class-performance-mlperf-inference,2026-10-28,MLPerf Inference v5.1
+NE2,Nebius,Nebius v5.1,"These two numbers also hold the first place for Nebius, among other ... Inference v5.1 submitters for this model on GB200 systems",true,"4-GPU GB200 entries in v5.1, Llama 3.1 405B",first place among v5.1 submitters on GB200,Nebius Offline 855.82. Azure Offline 847.23. Nebius Server 596.11. Azure Server 584.01,CONSISTENT,The group is the 4-GPU entries. The 72-GPU NVIDIA entries are higher.,false,Nebius: Nebius proves bare-metal-class performance for AI inference workloads in MLPerf Inference v5.1,https://nebius.com/blog/posts/bare-metal-class-performance-mlperf-inference,2026-10-28,MLPerf Inference v5.1
+NE3,Nebius,Nebius v5.1,"one host with 4x Blackwell GPUs achieved a 54.8% performance gain compared to an H200 host powered by 8x GPUs, while normalized per-chip comparison showed a 3.1x performance increase",true,"Nebius GB200 4-GPU entry against Nebius H200 8-GPU entry, Llama 3.1 405B Offline","+54.8% per host, 3.1x per chip","855.82 against 552.75 (+54.8%), per chip 3.10x",CONSISTENT,,false,Nebius: Nebius proves bare-metal-class performance for AI inference workloads in MLPerf Inference v5.1,https://nebius.com/blog/posts/bare-metal-class-performance-mlperf-inference,2026-10-28,MLPerf Inference v5.1
+NE4,Nebius,Nebius v5.1,"Running Llama 3.1 405B inference on NVIDIA HGX B200, we recorded 1,660 tokens/s in offline mode and 1,280 tokens/s in server scenario",true,"Nebius B200 8-GPU entry 5.1-0073, Llama 3.1 405B","1,660 and 1,280","1,659.5 and 1,279.5",CONSISTENT,Rounded to the nearest ten.,false,Nebius: Nebius proves bare-metal-class performance for AI inference workloads in MLPerf Inference v5.1,https://nebius.com/blog/posts/bare-metal-class-performance-mlperf-inference,2026-10-28,MLPerf Inference v5.1
+NE5,Nebius,Nebius v5.1,we see 3x and 4.3x performance increases in offline and server scenarios respectively,true,"Nebius B200 entry against Nebius H200 entry, Llama 3.1 405B","3x Offline, 4.3x Server","3.00x Offline, 4.33x Server",CONSISTENT,,false,Nebius: Nebius proves bare-metal-class performance for AI inference workloads in MLPerf Inference v5.1,https://nebius.com/blog/posts/bare-metal-class-performance-mlperf-inference,2026-10-28,MLPerf Inference v5.1
+NE6,Nebius,Nebius v5.1,"The B200 system outperforms the H200 3x and 2.9x times , by delivering 101,611 and 101,246 tokens/s in server and offline modes, respectively",true,"Nebius B200 and H200 entries, Llama 2 70B 99","3x Server, 2.9x Offline, 101,611 and 101,246","Server 101,611 (2.99x), Offline 101,246 (2.91x)",CONSISTENT,,false,Nebius: Nebius proves bare-metal-class performance for AI inference workloads in MLPerf Inference v5.1,https://nebius.com/blog/posts/bare-metal-class-performance-mlperf-inference,2026-10-28,MLPerf Inference v5.1
+NE8,Nebius,Nebius v5.1,Llama 70B inference on HGX H200 systems gives 11% performance increase compared to the best results achieved on H100-powered systems from the MLPerf ... Inference v5.0 submission,true,"Nebius H200 8-GPU entry against the best 8-GPU H100 entry in v5.0, Llama 2 70B 99",+11%,"Offline 34,812 against 31,307 (+11.2%). Server 34,029 against 31,106 (+9.4%)",CONSISTENT,The 11% reproduces for Offline. Server is 9.4%.,false,Nebius: Nebius proves bare-metal-class performance for AI inference workloads in MLPerf Inference v5.1,https://nebius.com/blog/posts/bare-metal-class-performance-mlperf-inference,2026-10-28,MLPerf Inference v5.1
+NE7,Nebius,Nebius v5.1,"For both models, Nebius demonstrated the best performance results in server mode , while in offline and interactive scenarios it shows parity with other submitters.",true,"Nebius 8-GPU B200 and H200 entries ranked within the 8-GPU entries of the same accelerator, v5.1","best in Server, parity in Offline and Interactive",llama2-70b-99 Server B200: rank 1 of 10. llama2-70b-99 Server H200: rank 2 of 10. llama2-70b-99 Offline B200: rank 8 of 10. llama2-70b-99 Offline H200: rank 6 of 11. llama2-70b-99 Interactive B200: rank 4 of 6. llama2-70b-99 Interactive H200: rank 1 of 6. llama3.1-405b Server B200: rank 1 of 13. llama3.1-405b Server H200: rank 2 of 3. llama3.1-405b Offline B200: rank 2 of 13. llama3.1-405b Offline H200: rank 1 of 3. llama3.1-405b Interactive B200: rank 1 of 10. llama3.1-405b Interactive H200: rank 1 of 1,CANNOT_TELL,Parity is not defined. Nebius is first in some Server rows and not in others.,false,Nebius: Nebius proves bare-metal-class performance for AI inference workloads in MLPerf Inference v5.1,https://nebius.com/blog/posts/bare-metal-class-performance-mlperf-inference,2026-10-28,MLPerf Inference v5.1
+LB8,Lambda,Lambda v6.1,Our Server result of 64.624 queries/s ranks second of four submissions,true,closed Qwen3-VL Server entries on four GB300 in the v6.1 file,second of four,"Lambda 64.624 ranks 1 of 4: Lambda 64.624, Nebius 64.620, Oracle 64.608, CoreWeave 64.132",ABOVE,"The file value rounds to 64.624. Lambda is first in the file, not second. ABOVE means the file shows more than the statement, here first place where the post says second.",false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB9,Lambda,Lambda v6.1,in a field where the entire spread from first (64.63 q/s) to fourth (64.608 q/s) is a few hundredths of a query per second,true,"closed Qwen3-VL Server entries on four GB300, first to fourth","64.63 to 64.608, a few hundredths","first 64.624 to fourth 64.132, spread 0.492 q/s. First to third spread 0.017",BELOW,"The fourth entry in the file is CoreWeave. The post figure 64.608 matches the Oracle entry, which is third in the file. Our rule: a few hundredths means under 0.1.",false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB1,Lambda,Lambda v6.1,our 4 NVIDIA Blackwell Ultra GPUs system delivers the leading Offline throughput on GPT-OSS 120B among all 4 Blackwell Ultra GPUs submissions,true,closed GPT-OSS-120B Offline entries on four NVIDIA GB300 in the v6.1 file,leading Offline result of four GB300 entries,"Lambda 65,511.9 tokens/s is first of 3, next Wiwynn 63,990.5",CONSISTENT,The post says Blackwell Ultra. The file lists these entries as NVIDIA GB300 with four accelerators.,false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB10,Lambda,Lambda v6.1,Our Offline result of 71.122 queries/s sits within 2.9% of the leading result (73.186 q/s).,true,Lambda Offline entry against the leading GB300 four-accelerator Qwen3-VL Offline entry,"71.122, within 2.9% of 73.186","Lambda 71.122. Leader Nebius 73.186. 2.82% below the leader, 2.90% measured against Lambda",CONSISTENT,The post gap of 2.9% matches the gap measured against Lambda. Measured against the leader it is 2.82%.,false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB11,Lambda,Lambda v6.1,"86.83% BFCL v4 accuracy; 1,007/1,007 replay turns completed",true,"Lambda open-division Kimi entry in the v6.1 file, accuracy field",86.83% BFCL v4 accuracy,"accuracy field 86.83%, mean latency 770.8 ms",CONSISTENT,"The replay turn count of 1,007 is not a file field. The accuracy field is the BFCL v4 figure.",false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB12,Lambda,Lambda v6.1,Mean per-turn latency came in at 770.8 ms,true,"Lambda open-division Kimi entry, SingleStream result in ms",770.8 ms,770.8 ms,CONSISTENT,The median of 361.0 ms and the time to first token figures are not file fields.,false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB13,Lambda,Lambda v6.1,"The comparison table in the post lists mean latency in ms and BFCL v4 accuracy for the Lambda open entry (770.8, 86.83%) and three anonymous closed edge entries: 1,466.0 and 87.94%, 2,616.2 and 86.43%, 2,758.0 and 86.13%, and 3,807.7 and 87.24%.",false,table rows for the four closed edge entries,"latency and accuracy for Vendors A, B and C",Thor 1466.0 ms 87.94% . DL345 2616.2 ms 86.43% . DL145 2758.0 ms 86.13% . Spark 3807.7 ms 87.24%,CONSISTENT,Vendor labels in the post are anonymous. Matched by system name. Vendor B maps to the two ProLiant entries and Vendor C to the DGX Spark entry.,false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB15,Lambda,Lambda v6.1,Dual Intel Xeon Platinum 8570 (56 cores),true,Lambda B200 entry host processor fields,"dual Xeon Platinum 8570, 56 cores","Intel(R) Xeon(R) Platinum 8570, 2 processors, 56 cores each",CONSISTENT,,false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB16,Lambda,Lambda v6.1,This submission marks Lambda's first MLPerf result for a vision-language model (VLM).,true,"Lambda entries in the v5.0, v5.1 and v6.0 files, models with VL in the name",first Lambda vision-language result,0 Lambda vision-language entries before v6.1,CONSISTENT,,false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB2,Lambda,Lambda v6.1,that same system posts an 8.85% (Server) and 8.79% (Offline) throughput improvement over v6.0 on identical hardware,true,"Lambda entry in v6.1 against Lambda_SIT entry 6.0-0063, GPT-OSS-120B","+8.85% Server, +8.79% Offline","+8.85% Server (58,195.8 vs 53,462.7), +8.79% Offline (65,511.9 vs 60,220.4)",CONSISTENT,The v6.0 entry lists NVIDIA GB300 with 4 accelerators. Whether the hardware is identical is not a file field.,false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB3,Lambda,Lambda v6.1,"The Offline result leads all 4 Blackwell Ultra GPUs submissions this round, besting second place by 2.38% and third place by 7.8%.",true,Lambda Offline entry against the second and third GB300 four-accelerator entries,"+2.38% over second, +7.8% over third","+2.38% over Wiwynn (63,990.5), +7.85% over Oracle (60,745.8)",CONSISTENT,The third place figure of 7.85% rounds to 7.8% as the post states.,false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB4,Lambda,Lambda v6.1,"On the Server scenario, our result lands within 3.9% of the fastest submission (60,466 tokens/s), a tight three-way field where the spread between first and third is under 4%.",true,Lambda Server entry against the fastest GB300 four-accelerator Server entry,"within 3.9% of 60,466. First to third under 4%","fastest Wiwynn 60,466.2. Lambda is 3.75% below it, or 3.90% if measured against Lambda. Lambda ranks 3 of 3 among the three highest GB300 entries",CONSISTENT,"The post gives 3.9%, which matches the gap measured against Lambda. Measured against the fastest it is 3.76%. Both are within the stated bound.",false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB5,Lambda,Lambda v6.1,The Offline result of 101.56 queries/s is 28.5% faster than the fastest submission from the previous round (79.04 q/s),true,Lambda B200 Qwen3-VL Offline entry against the fastest v6.0 closed Qwen3-VL Offline entry,"101.56 q/s, +28.5% over 79.04","101.56 vs 79.04, +28.50%",CONSISTENT,,false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB6,Lambda,Lambda v6.1,the Server result of 69.41 queries/s edges out the prior round's best (67.86 q/s) by 2.29%,true,Lambda B200 Qwen3-VL Server entry against the fastest v6.0 closed Qwen3-VL Server entry,"69.41 q/s, +2.29% over 67.86","69.42 vs 67.86, +2.29%",CONSISTENT,"The file value is 69.4154, which rounds to 69.42 at two places. The post prints 69.41. The gap is the fourth decimal.",false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB7,Lambda,Lambda v6.1,"On NVIDIA HGX B200, ours is the only Qwen3-VL-235B-A22B-Instruct submission on this configuration.",true,closed Qwen3-VL entries on eight B200 accelerators in the v6.1 file,only submission on eight B200,1 organization: Lambda,CONSISTENT,,false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB14,Lambda,Lambda v6.1,ours is the only agentic inference submission on datacenter hardware,true,"all qwen3.6-27b SingleStream entries in the v6.1 file, SystemType field",only agentic submission on datacenter hardware,"6 entries, SystemType values edge, Lambda entry SystemType edge",CANNOT_TELL,"The file lists every agentic entry including Lambda as edge. Datacenter hardware is a description of the machine, not a file field. 8 B200 accelerators is the largest system among the agentic entries.",false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LB17,Lambda,Lambda v6.1,The post states the four Blackwell Ultra GPUs deliver 1.54x Offline and 1.42x Server speedups over a four-GPU baseline estimated by halving Lambda's HGX B200 closed GPT-OSS 120B result.,false,speedup of the four-GPU entry over half of Lambda HGX B200 closed GPT-OSS-120B,1.54 Offline and 1.42 Server,"Lambda closed GPT-OSS-120B rows on B200 in the v6.1 file: 0. Implied B200 baselines would be 85,080 Offline and 81,966 Server for eight GPUs",CANNOT_TELL,The post does not give the B200 GPT-OSS-120B figure and the v6.1 file has no Lambda B200 row for that model.,false,"Lambda: MLPerf Inference v6.1: pioneering agent, VLM benchmarks (post dated 16 September 2026, updated 22 October 2026)",https://lambda.ai/blog/mlperf-inference-v6.1,2026-10-28,MLPerf Inference v6.1
+LN1,Lenovo,Lenovo v5.1,earning Top 3 rankings and 1st place in fifteen categories,true,"first-place counts for the SR680a V3, SR780a V3 and SE100 entries under the rank rule in the notes",1st place in fifteen categories,"SR680a 2, SR780a 1, SE100 4 first places, total 7",BELOW,"Our rank rule: datacenter entries ranked among the closed 8-accelerator B200 entries for the same model and scenario, which reproduces the field sizes in the list. Edge entries ranked among all closed edge entries for the same model and scenario. Singlestream and multistream rank by lower latency.",false,"Lenovo: Unleashing AI Excellence: Lenovo ThinkSystem Dominates MLPerf v5.1 with Exceptional Inference Performance (Lenovo Press LP2330, published 11 November 2025)",https://lenovopress.lenovo.com/lp2330-lenovo-thinksystem-dominates-mlperf-v51-with-exceptional-inference-performance,2026-10-28,MLPerf Inference v5.1
+LN4,Lenovo,Lenovo v5.1,The Table 1 row for ThinkEdge SE100 in the post lists twelve categories submitted with twelve first place finishes and no second or third place finishes.,false,"SE100 closed edge entry 5.1-0064, rank among closed edge entries for each model and scenario","12 categories, 12 first, 0 second, 0 third","12 categories, 4 first, 6 second, 2 third. Fields: resnet SingleStream 2/7, resnet MultiStream 2/7, resnet Offline 2/7, retinanet SingleStream 2/7, retinanet MultiStream 3/7, retinanet Offline 3/7, 3d-unet-99 SingleStream 2/2, 3d-unet-99 Offline 2/2, 3d-unet-99.9 SingleStream 1/1, 3d-unet-99.9 Offline 1/1, stable-diffusion-xl SingleStream 1/1, stable-diffusion-xl Offline 1/1",BELOW,"The list in the post says 1/1 and no competitors for all twelve. The file has other closed edge entries in eight of the twelve, from GATEOverflow and HPE, and one other entry in two more.",false,"Lenovo: Unleashing AI Excellence: Lenovo ThinkSystem Dominates MLPerf v5.1 with Exceptional Inference Performance (Lenovo Press LP2330, published 11 November 2025)",https://lenovopress.lenovo.com/lp2330-lenovo-thinksystem-dominates-mlperf-v51-with-exceptional-inference-performance,2026-10-28,MLPerf Inference v5.1
+LN6,Lenovo,Lenovo v5.1,"The Table 1 row for ThinkSystem SR780a V3 (Intel) in the post lists nine categories submitted with one first place finish, two second place finishes and two third place finishes.",false,"SR780a V3 entry 5.1-0062, ranks within closed 8-accelerator B200 entries","9 categories, 1 first, 2 second, 2 third","9 categories, 1 first, 2 second, 0 third",BELOW,"The ranks listed under the table in the post also give no third place for this system: 4, 1, 4, 6, 2, 9, 2, 4 and 5.",false,"Lenovo: Unleashing AI Excellence: Lenovo ThinkSystem Dominates MLPerf v5.1 with Exceptional Inference Performance (Lenovo Press LP2330, published 11 November 2025)",https://lenovopress.lenovo.com/lp2330-lenovo-thinksystem-dominates-mlperf-v51-with-exceptional-inference-performance,2026-10-28,MLPerf Inference v5.1
+LN8,Lenovo,Lenovo v5.1,"Under the SR780a V3 heading the post lists nine ranks: 1-8b server 4/9 and offline 1/9, stable-diffusion-xl server 4/8 and offline 6/9, 1-405b server 2/13, offline 9/13 and interactive 2/10, rgat 4/5, and whisper 5/14.",false,SR780a V3 entry 5.1-0062 ranks within closed 8-accelerator B200 entries,"4/9, 1/9, 4/8, 6/9, 2/13, 9/13, 2/10, 4/5, 5/14","llama3.1-8b Server 4/9, llama3.1-8b Offline 1/9, stable-diffusion-xl Server 4/8, stable-diffusion-xl Offline 6/8, llama3.1-405b Server 2/13, llama3.1-405b Offline 9/13, llama3.1-405b Interactive 2/10, rgat Offline 4/5, whisper Offline 5/14",BELOW,"Differences from the post: stable-diffusion-xl Offline stated 6/9, file 6/8. The rank itself matches in each case, the field size differs by one.",false,"Lenovo: Unleashing AI Excellence: Lenovo ThinkSystem Dominates MLPerf v5.1 with Exceptional Inference Performance (Lenovo Press LP2330, published 11 November 2025)",https://lenovopress.lenovo.com/lp2330-lenovo-thinksystem-dominates-mlperf-v51-with-exceptional-inference-performance,2026-10-28,MLPerf Inference v5.1
+LN9,Lenovo,Lenovo v5.1,"Under the ThinkEdge SE100 heading the post lists twelve entries for resnet50, retinanet, 3d-unet, 3d-unet-99.9 and stable-diffusion-xl, each marked 1/1 with no competitors.",false,closed edge entries for the same model and scenario as each SE100 entry,12 categories with no competitors,"8 of 12 have other closed edge entries. Field sizes: resnet SingleStream 7, resnet MultiStream 7, resnet Offline 7, retinanet SingleStream 7, retinanet MultiStream 7, retinanet Offline 7, 3d-unet-99 SingleStream 2, 3d-unet-99 Offline 2, 3d-unet-99.9 SingleStream 1, 3d-unet-99.9 Offline 1, stable-diffusion-xl SingleStream 1, stable-diffusion-xl Offline 1",BELOW,Rank 1 does not hold in every one of the twelve.,false,"Lenovo: Unleashing AI Excellence: Lenovo ThinkSystem Dominates MLPerf v5.1 with Exceptional Inference Performance (Lenovo Press LP2330, published 11 November 2025)",https://lenovopress.lenovo.com/lp2330-lenovo-thinksystem-dominates-mlperf-v51-with-exceptional-inference-performance,2026-10-28,MLPerf Inference v5.1
+LN3,Lenovo,Lenovo v5.1,"AMD, Azure, Broadcom, Cisco, Dell, Google, Intel and HPE",true,Submitter values in the v5.1 file containing each name,eight named participants,8 of 8 found as submitter names or parts of combined names. Missing: none,CONSISTENT,"Broadcom appears only in combined names (Broadcom_Supermicro, Dell_Broadcom).",false,"Lenovo: Unleashing AI Excellence: Lenovo ThinkSystem Dominates MLPerf v5.1 with Exceptional Inference Performance (Lenovo Press LP2330, published 11 November 2025)",https://lenovopress.lenovo.com/lp2330-lenovo-thinksystem-dominates-mlperf-v51-with-exceptional-inference-performance,2026-10-28,MLPerf Inference v5.1
+LN5,Lenovo,Lenovo v5.1,"The Table 1 row for ThinkSystem SR680a V3 in the post lists five categories submitted with two first place finishes, no second place finishes and one third place finish.",false,"SR680a V3 entry 5.1-0061, ranks within closed 8-accelerator B200 entries","5 categories, 2 first, 0 second, 1 third","5 categories, 2 first, 0 second, 1 third",CONSISTENT,,false,"Lenovo: Unleashing AI Excellence: Lenovo ThinkSystem Dominates MLPerf v5.1 with Exceptional Inference Performance (Lenovo Press LP2330, published 11 November 2025)",https://lenovopress.lenovo.com/lp2330-lenovo-thinksystem-dominates-mlperf-v51-with-exceptional-inference-performance,2026-10-28,MLPerf Inference v5.1
+LN7,Lenovo,Lenovo v5.1,"Under the SR680a V3 heading the post lists five ranks: llama2-70b-99 server 7/10 and offline 1/10, llama2-70b-99.9 server 3/11 and offline 1/11, and whisper 7/14.",false,SR680a V3 entry 5.1-0061 ranks within closed 8-accelerator B200 entries,"7/10, 1/10, 3/11, 1/11, 7/14","llama2-70b-99 Server 7/10, llama2-70b-99 Offline 1/10, llama2-70b-99.9 Server 3/11, llama2-70b-99.9 Offline 1/11, whisper Offline 7/14",CONSISTENT,,false,"Lenovo: Unleashing AI Excellence: Lenovo ThinkSystem Dominates MLPerf v5.1 with Exceptional Inference Performance (Lenovo Press LP2330, published 11 November 2025)",https://lenovopress.lenovo.com/lp2330-lenovo-thinksystem-dominates-mlperf-v51-with-exceptional-inference-performance,2026-10-28,MLPerf Inference v5.1
+LN2,Lenovo,Lenovo v5.1,twenty-seven participating organizations,true,distinct Submitter values in the v5.1 file,twenty-seven participating organizations,"32 distinct submitter names in the file, 29 with a closed entry",CANNOT_TELL,"Submitter names in the file include combined names such as Dell_Broadcom and Supermicro_MangoBoost, so the count of organizations depends on how those are split. The count of organizations MLCommons announced is not a file field.",false,"Lenovo: Unleashing AI Excellence: Lenovo ThinkSystem Dominates MLPerf v5.1 with Exceptional Inference Performance (Lenovo Press LP2330, published 11 November 2025)",https://lenovopress.lenovo.com/lp2330-lenovo-thinksystem-dominates-mlperf-v51-with-exceptional-inference-performance,2026-10-28,MLPerf Inference v5.1
+NV1,Nebius,Nebius v6.1,Nebius achieved five first-place results across 20 available category entries,true,"Nebius closed available v6.1 entries ranked among closed available entries with the same model, scenario, accelerator and accelerator count","five first-place results, 20 category entries",5 first places across 20 ranked-table entries,CONSISTENT,Our rank rule: entries with the same accelerator model and count. It reproduces every rank shown in the post table.,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV10,Nebius,Nebius v6.1,"we ranked first in both the server and offline scenarios for DeepSeek R1, at 603,023 and 689,961 tokens/s respectively",true,Nebius 72 GPU GB300 DeepSeek R1 entries ranked among 72 GPU GB300 entries,"first in Server and Offline, 603,023 and 689,961","Server 603,023 rank 1 of 4, Offline 689,961 rank 1 of 4",CONSISTENT,,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV11,Nebius,Nebius v6.1,The same full-rack system sustained over 1.1 million tokens/s on gpt-oss 120B in both scenarios,true,Nebius 72 GPU GB300 gpt-oss 120B entries,over 1.1 million tokens/s in both scenarios,"Server 1,122,490, Offline 1,186,750",CONSISTENT,,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV12,Nebius,Nebius v6.1,"moving from the eight-GPU GB300 NVL72 configuration to the NVIDIA GB300 NV72 rack, a 9x increase in GPUs, it delivered 8.8x higher server throughput and 9.0x higher offline throughput",true,"Nebius 72 GPU entry over Nebius 8 GPU entry, gpt-oss 120B","9x GPUs, 8.8x Server, 9.0x Offline","9x GPUs, 8.77x Server (1,122,490 over 127,921), 8.97x Offline (1,186,750 over 132,236)",CONSISTENT,,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV13,Nebius,Nebius v6.1,Only two submitters ran that eight-GPU configuration.,true,closed 8 GPU GB300 entries in the v6.1 file,two submitters,"gpt-oss 120B: NVIDIA and Nebius. Any model: NVIDIA, Nebius and Oracle",CONSISTENT,True for gpt-oss 120B. Oracle also has eight-GPU GB300 entries for other models.,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV14,Nebius,Nebius v6.1,Nebius posted the top offline result for gpt-oss 120B and came within 0.2% of the top server result,true,Nebius 8 GPU GB300 entries against the other 8 GPU GB300 gpt-oss 120B entry,"top Offline, within 0.2% of top Server","Offline 132,236 top is Nebius. Server top NVIDIA 128,173, Nebius 0.20% below",CONSISTENT,"The post sentence follows the eight-GPU comparison, so the top is read within that group.",false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV15,Nebius,Nebius v6.1,"raised server throughput from 55,869 to 63,909 tokens/s on identical eight-GPU configurations",true,Nebius B200 and B300 DeepSeek R1 Server entries,"55,869 to 63,909, eight GPUs","55,869 to 63,909 (+14.4%), 8 and 8 accelerators",CONSISTENT,"The entries differ in accelerator model, which is the comparison being made.",false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV16,Nebius,Nebius v6.1,"from 89,856 to 108,124 tokens/s",true,Nebius B200 and B300 gpt-oss 120B Server entries,"89,856 to 108,124","89,856 to 108,124 (+20.3%)",CONSISTENT,,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV17,Nebius,Nebius v6.1,That node posted the fastest offline result among all 8x NVIDIA RTX PRO 6000 GPU submissions for gpt-oss 120B and came within 0.3% of the best server result,true,Nebius RTX PRO 6000 8 GPU entries against all 8 GPU RTX PRO 6000 gpt-oss 120B entries,"fastest Offline, within 0.3% of best Server","Offline top Nebius 15,738.5. Server top Cisco 14,817.4, Nebius 0.22% below",CONSISTENT,,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV2,Nebius,Nebius v6.1,"The results table in the post lists twenty values for seven Nebius systems across DeepSeek R1, Qwen3-VL 235B and gpt-oss 120B in Server and Offline, plus two preview values for the Vera Rubin system.",false,"Nebius entries in the v6.1 file, result field rounded to the digits printed",twenty table values,20 of 20 values match the file to the digits printed,CONSISTENT,,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV3,Nebius,Nebius v6.1,"Beside seventeen of the twenty values the table prints a rank from 1st to 5th, and three values carry a dash in place of a rank. Each value is followed by a reference number to an MLCommons entry.",false,"rank of each Nebius entry in its group of same model, scenario, accelerator and count",seventeen printed ranks,B200 n1 deepse Ser 2/2. B200 n1 deepse Off 2/2. B200 n1 gpt-os Ser 2/2. B200 n1 gpt-os Off 2/2. B300 n1 deepse Ser 5/13. B300 n1 deepse Off 10/13. B300 n1 qwen3- Ser 2/6. B300 n1 qwen3- Off 3/7. B300 n1 gpt-os Ser 6/13. B300 n1 gpt-os Off 8/13. RTX PRO 6000 gpt-os Ser 2/3. RTX PRO 6000 gpt-os Off 1/3. 4x GB300 qwen3- Ser 2/4. 4x GB300 qwen3- Off 1/4. 8x GB300 gpt-os Ser 2/2. 8x GB300 gpt-os Off 1/2. 72x GB300 deepse Ser 1/4. 72x GB300 deepse Off 1/4. 72x GB300 gpt-os Ser 2/4. 72x GB300 gpt-os Off 2/4,CONSISTENT,"Three cells carry a dash instead of a rank (B300 DeepSeek R1 Offline, B300 gpt-oss 120B Server and Offline). The file ranks for those are 10 of 13, 6 of 13 and 8 of 13.",false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV4,Nebius,Nebius v6.1,Nebius was one of only two submitters in this round with results on Vera Rubin-based hardware.,true,v6.1 file entries whose accelerator name contains Rubin or VR200,one of two submitters,organizations: NVIDIA and Nebius,CONSISTENT,,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV5,Nebius,Nebius v6.1,Seven system configurations span a wide range of inference hardware,true,distinct Nebius SystemName values in the v6.1 file,seven system configurations,7 system names,CONSISTENT,,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV6,Nebius,Nebius v6.1,"We benchmarked three models in this round: DeepSeek R1, Qwen3-VL 235B, and gpt-oss 120B, in both the server and offline scenarios.",true,distinct Nebius models and scenarios in the v6.1 file,"three models, server and offline","3 models deepseek-r1, gpt-oss-120b and qwen3-vl-235b-a22b. Each has Server and Offline: yes",CONSISTENT,,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV7,Nebius,Nebius v6.1,Nebius ran DeepSeek R1 on 36 GPUs across nine nodes on a Nebius system with NVIDIA Vera Rubin NVL72 in the preview category.,true,Nebius VR200 entry fields,"36 GPUs, nine nodes, preview","36 accelerators, 9 nodes, preview",CONSISTENT,,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV8,Nebius,Nebius v6.1,"On that basis, Nebius posted the leading server result at 16,427 tokens/s per GPU",true,"preview DeepSeek R1 Server entries on Vera Rubin, result divided by accelerator count","leading per-GPU Server result, 16,427 tokens/s","Nebius 16,427, NVIDIA 16,332",CONSISTENT,Nebius is the higher of the two preview entries.,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+NV9,Nebius,Nebius v6.1,72 NVIDIA Blackwell Ultra GPUs across 18 nodes served large-scale inference workloads.,true,Nebius 72 GPU entry fields,"72 GPUs, 18 nodes","72 accelerators, 18 nodes",CONSISTENT,,false,Nebius: MLPerf Inference v6.1 Results: NVIDIA Vera Rubin NVL72,https://nebius.com/blog/posts/mlperf-inference-v6-1-results,2026-10-28,MLPerf Inference v6.1
+RO1,Red Hat,Red Hat v5.0,Red Hat OpenShift running on their dual GPU GH200 Grace Hopper Superchip 144GB server,true,v5.0 file entries whose operating_system field names OpenShift,OpenShift on a dual GH200 144GB server,"2 entries 5.0-5663 and 5.0-5664: Supermicro, open division, 2 accelerators, ARS-221GL-NHIR (2x GH200-144GB_aarch64, TensorRT), operating_system Red Hat 9.4 + OpenShift 4.15",CONSISTENT,"The file lists the submitter as Supermicro, in the open division, under entries 5.0-5663 and 5.0-5664 (llama2-70b-99 and 99.9).",false,Red Hat: MLPerf Inference v5.0 results with Supermicro's GH200 Grace Hopper Superchip-based Server and Red Hat OpenShift (post dated 18 June 2025),https://www.redhat.com/en/blog/mlperf-inference-v50-results,2026-10-28,MLPerf Inference v5.0
+RO2,Red Hat,Red Hat v5.0,This was the first time anyone has submitted an MLPerf result with OpenShift on GH200.,true,"OpenShift entries in the v5.0 file, and system description files with GH200 or Grace in the name in the v3.1, v4.0 and v4.1 repositories",first OpenShift on GH200 submission,"2 OpenShift entries in v5.0 (both this system). In v3.1, v4.0 and v4.1: 12 GH200 or Grace system files, 0 naming OpenShift",CONSISTENT,"Earlier rounds have no summary file, so the system description files in each repository were read. The v3.1 and v4.0 file lists from the repository API were cut short at the end of the list. Red Hat submitted OpenShift once before, on an L40S system in v4.1.",false,Red Hat: MLPerf Inference v5.0 results with Supermicro's GH200 Grace Hopper Superchip-based Server and Red Hat OpenShift (post dated 18 June 2025),https://www.redhat.com/en/blog/mlperf-inference-v50-results,2026-10-28,MLPerf Inference v5.0
+RO3,Red Hat,Red Hat v5.0,"there were 17 organizations who submitted Llama 2 70B results, making it the most popular model in this round",true,closed datacenter v5.0 entries: distinct submitters per model,"17 organizations, most popular model","llama2-70b-99 has 17 submitters, next llama2-70b-99.9 16. 19 submitters if open and interactive entries are counted",CONSISTENT,The count of 17 matches closed datacenter llama2-70b-99. Counting every llama2-70b variant and division gives 19.,false,Red Hat: MLPerf Inference v5.0 results with Supermicro's GH200 Grace Hopper Superchip-based Server and Red Hat OpenShift (post dated 18 June 2025),https://www.redhat.com/en/blog/mlperf-inference-v50-results,2026-10-28,MLPerf Inference v5.0
+RO4,Red Hat,Red Hat v5.0,The Supermicro MLPerf v5.0 dual GPU GH200 server submission ran OpenShift 4.15 and NVIDIA TRT-LLM for the server stack.,true,v5.0 entry 5.0-5663 operating_system and Software fields,OpenShift 4.15 and TRT-LLM,"operating_system Red Hat 9.4 + OpenShift 4.15. Software TensorRT 10.8, CUDA 12.8",CONSISTENT,The file names TensorRT 10.8 and CUDA 12.8 and does not use the name TRT-LLM.,false,Red Hat: MLPerf Inference v5.0 results with Supermicro's GH200 Grace Hopper Superchip-based Server and Red Hat OpenShift (post dated 18 June 2025),https://www.redhat.com/en/blog/mlperf-inference-v50-results,2026-10-28,MLPerf Inference v5.0
+RO5,Red Hat,Red Hat v5.0,TRT-LLM uses post-training quantization to quantize llama2-70b to FP8 precision (8-bit floating point).,true,v5.0 entry 5.0-5663 weight_data_types field,FP8 weights,weight_data_types fp8,CONSISTENT,,false,Red Hat: MLPerf Inference v5.0 results with Supermicro's GH200 Grace Hopper Superchip-based Server and Red Hat OpenShift (post dated 18 June 2025),https://www.redhat.com/en/blog/mlperf-inference-v50-results,2026-10-28,MLPerf Inference v5.0
+RO6,Red Hat,Red Hat v5.0,The chart below shows that OpenShift added less than 2% overhead in these four scenarios for Llama2-70b.,true,OpenShift entries in the v5.0 file against any RHEL 9.4 bare-metal entry for the same system,less than 2% overhead in four scenarios,"4 OpenShift rows (Offline 8,231.9, Server 7,538.8, same for 99.9). RHEL 9.4 rows for this system in the file: 0",CANNOT_TELL,The overhead is a comparison with a RHEL 9.4 run that is not an entry in the result file. The chart is not a file field.,false,Red Hat: MLPerf Inference v5.0 results with Supermicro's GH200 Grace Hopper Superchip-based Server and Red Hat OpenShift (post dated 18 June 2025),https://www.redhat.com/en/blog/mlperf-inference-v50-results,2026-10-28,MLPerf Inference v5.0
+RR12,Red Hat,Red Hat v6.0,The table in the post lists Red Hat Qwen3-VL-235B-A22B on 8x H200 at 18.02 samples/sec Offline and 11.05 samples/sec Server.,false,"Red Hat Qwen3-VL H200 entry 6.0-0095, Offline","18.02 samples/s Offline, 11.05 Server","Offline 18.218, Server 11.05",ABOVE,"The text under the table gives 18.22, which matches the file. The table prints 18.02.",false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR6,Red Hat,Red Hat v6.0,"8x H200 offline: 36,396 tokens per second , the leading H200 result, 13% faster than the next closest submission",true,Red Hat H200 Whisper Offline entry against the next H200 Whisper Offline entry,"36,396 tokens/s, +13%","36,395.7 against 31,969.2 (Dell, NVIDIA H200-NVL-141GB), +13.85%",ABOVE,The gain rounds to 14%. The next entry uses the H200 NVL card.,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR13,Red Hat,Red Hat v6.0,"The table in the post lists llama-2-70b on 8x MI350X with Supermicro at 91,933.10 Offline and 89,019.65 Server.",false,Supermicro_RedHat llama2-70b-99 entry 6.0-0100,"91,933.10 Offline, 89,019.65 Server","80,478.1 Offline, 76,393.1 Server. The text under the table gives 80,478 and 76,393",BELOW,"No v6.0 entry has either table figure. The text under the table gives 80,478 and 76,393, which match the file.",false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR1,Red Hat,Red Hat v6.0,"Our submission includes four AI workloads (Whisper-Large-v3, GPT-OSS-120B, Qwen3-VL-235B-A22B, and Llama-2-70b) on NVIDIA (H200, B200, L40S) and AMD (MI350X) GPUs",true,v6.0 entries from RedHat and Supermicro_RedHat,"four workloads on H200, B200, L40S and MI350X","models gpt-oss-120b, llama2-70b-99, qwen3-vl-235b-a22b and whisper. Accelerators B200, H200, L40S and MI350X",CONSISTENT,Llama-2-70b appears in the file as llama2-70b-99.,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR10,Red Hat,Red Hat v6.0,"2x L40S offline: 3,647 tokens per second",true,Red Hat Whisper L40S entry 6.0-0090,"3,647 tokens/s","3,646.91",CONSISTENT,,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR11,Red Hat,Red Hat v6.0,"Offline: 80,478 tokens per second ... Server: 76,393 tokens per second",true,Supermicro_RedHat llama2-70b-99 entry 6.0-0100,"80,478 and 76,393","80,478.1 and 76,393.1",CONSISTENT,,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR14,Red Hat,Red Hat v6.0,our results align with the expected ~80% performance ratio relative to the liquid-cooled MI355X,true,Supermicro_RedHat MI350X entry against the AMD eight-GPU MI355X llama2-70b-99 entry 6.0-0002,about 80% of the MI355X,"Offline 77.8% (80,478 of 103,480), Server 76.2% (76,393 of 100,282)",CONSISTENT,"Our rule: about 80% is read to the nearest ten, so 75 to 85%.",false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR18,Red Hat,Red Hat v6.0,reduced P99 time-to-first-token (TTFT) from 3.4 seconds to 2.1 seconds (~38% improvement),true,arithmetic on the two times in the post,~38% improvement,38.2% lower,CONSISTENT,Not a result file field. Checked from the post's own two numbers.,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR2,Red Hat,Red Hat v6.0,the highest offline throughput on B200 for GPT-OSS-120B,true,closed v6.0 GPT-OSS-120B Offline entries on B200,highest Offline on B200,"RedHat 93,070.7. Next Nebius 85,921.2. Top is RedHat",CONSISTENT,,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR20,Red Hat,Red Hat v6.0,Top in the leaderboard compared to all results,true,"closed v6.0 Qwen3-VL entries, all accelerators",top Offline and top Server,Offline top RedHat 79.04. Server top RedHat 67.86,CONSISTENT,,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR3,Red Hat,Red Hat v6.0,the leading H200 result on Whisper,true,closed v6.0 Whisper Offline entries on H200,leading H200 result,"RedHat 36,395.7. Next Dell 31,969.2",CONSISTENT,,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR4,Red Hat,Red Hat v6.0,"the top B200 submission on Qwen3-VL, which exceeded the top B300 performer by 50% in the Server scenario",true,closed v6.0 Qwen3-VL Server entries on B200 and on eight B300,"top B200, +50% over top B300 Server",RedHat 67.86 is top B200 of 1. Top eight-GPU B300 Nebius 45.15. +50.3%,CONSISTENT,The top B300 figure is read among the eight-GPU B300 entries. The four-GPU GB300 entries are lower.,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR5,Red Hat,Red Hat v6.0,"8x B200 offline: 93,071 tokens per second , 8% higher than the next comparable B200 submission",true,Red Hat B200 Offline entry against the next B200 Offline entry,"93,071 tokens/s, +8%","93,070.7 against 85,921.2, +8.32%",CONSISTENT,,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR7,Red Hat,Red Hat v6.0,the only H200 submission for this model in this round ... the only MI350X submission for this model this round ... the first and only L40S submission for Whisper in MLPerf Inference v6.0,true,v6.0 closed entries by model and accelerator. MI350X GPT-OSS-120B across both divisions,"only H200 for Qwen3-VL and GPT-OSS-120B, only MI350X for GPT-OSS-120B, only L40S for Whisper","H200 Qwen3-VL 1, H200 GPT-OSS-120B 1, MI350X GPT-OSS-120B 1 (division open), L40S Whisper 1",CONSISTENT,"The MI350X GPT-OSS-120B entry is in the open division, so there is no closed MI350X entry for that model.",false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR8,Red Hat,Red Hat v6.0,"8x B200 server: 67.86 samples per second ... 8x B200 offline: 79.04 samples per second ... 8x H200: 18.22 samples per second (offline), 11.05 samples per second (server)",true,Red Hat Qwen3-VL entries 6.0-0094 and 6.0-0095,"67.86, 79.04, 18.22 and 11.05","67.86, 79.04, 18.22, 11.05",CONSISTENT,,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR9,Red Hat,Red Hat v6.0,"8x B200 server: 71,588 tokens per second ... 8x H200: 28,680 tokens per second (offline), 24,103 tokens per second (server) ... 8XMI350X: 64,293 tokens per second (offline), 58,373 tokens per second (server)",true,"Red Hat GPT-OSS-120B entries 6.0-0091, 6.0-0092 and 6.0-0112","71,588, 28,680, 24,103, 64,293 and 58,373","71,588.1, 28,680.0, 24,103.2, 64,293.3, 58,373.3",CONSISTENT,,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR15,Red Hat,Red Hat v6.0,This was Red Hat's first MLPerf submission on AMD hardware.,true,"RedHat and Supermicro_RedHat entries on AMD accelerators in the v5.0, v5.1 and v6.0 files",first submission on AMD,"AMD entries by RedHat in v5.0: 0, v5.1: 0",CANNOT_TELL,None before v6.0 in the files used here. Rounds before v5.0 are not in them.,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR16,Red Hat,Red Hat v6.0,This makes our GPT-OSS-120B submission the first for this model on Kubernetes infrastructure.,true,GPT-OSS-120B entries in the v6.0 file,first on Kubernetes,30 GPT-OSS-120B entries in v6.0. Kubernetes is not a file field,CANNOT_TELL,"The file names the software stack, not the orchestrator.",false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR17,Red Hat,Red Hat v6.0,"We were able to achieve 20% gain in offline performance, and 24% gain in server performance post the submission date.",true,post-submission results are not result file entries,20% Offline and 24% Server gain after submission,no post-submission entry in the files,CANNOT_TELL,,false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+RR19,Red Hat,Red Hat v6.0,20 submitters in MLPerf v6.0 used vLLM (up from 5 in v5.1),true,"submitters whose v5.1 and v6.0 records name vLLM in Software, Platform, System or Notes","20 in v6.0, 5 in v5.1","v6.0 10 submitters, v5.1 13 submitters name vLLM",CANNOT_TELL,"Our rule: a record that names vLLM in any of the four text fields. The file shows two directions, so the row reads CANNOT_TELL: the v6.0 count is below the stated 20 and the v5.1 count is above the stated 5. A submitter that uses vLLM without naming it is not counted.",false,"Red Hat: Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B (post dated 1 April 2026)",https://www.redhat.com/en/blog/red-hat-ai-tops-mlperf-inference-v60-vllm-qwen3-vl-whisper-and-gpt-oss-120b,2026-10-28,MLPerf Inference v6.0
+CV1,CoreWeave,CoreWeave v6.1,"CoreWeave participated in the Datacenter Closed division’s Available category, submitting results across four NVIDIA platforms and four model families that are in use today.",true,CoreWeave entries in the v6.1 file,"closed, available, four platforms, four model families","division and availability: closed available. Platforms B200, B300, GB200 and GB300. Models deepseek-r1, gpt-oss-120b, llama2-70b-99.9 and qwen3-vl-235b-a22b",CONSISTENT,,false,CoreWeave: CoreWeave Leads Cloud Providers in MLPerf Inference v6.1 Performance with NVIDIA Blackwell Ultra,https://coreweave.com/blog/coreweave-leads-cloud-providers-in-mlperf-r-inference-v6-1-performance-with-nvidia-blackwell-ultra,2026-10-28,MLPerf Inference v6.1
+CV10,CoreWeave,CoreWeave v6.1,"In the offline scenario, we achieved the only result above one million tokens per second among cloud provider submissions using NVIDIA GB300 NVL72.",true,closed available v6.1 Llama 2 70B Offline entries from cloud providers on 72 GB300,"only one above 1,000,000",1 above one million: CoreWeave,CONSISTENT,CoreWeave is the only cloud provider entry in this group in the file.,false,CoreWeave: CoreWeave Leads Cloud Providers in MLPerf Inference v6.1 Performance with NVIDIA Blackwell Ultra,https://coreweave.com/blog/coreweave-leads-cloud-providers-in-mlperf-r-inference-v6-1-performance-with-nvidia-blackwell-ultra,2026-10-28,MLPerf Inference v6.1
+CV11,CoreWeave,CoreWeave v6.1,"highest aggregate throughput of any NVIDIA-based system, on a single GB300 NVL72 rack",true,closed available v6.1 GPT-OSS-120B Offline entries on NVIDIA accelerators,"highest aggregate of any NVIDIA-based system, single GB300 rack","all NVIDIA entries top three: CoreWeave 1,197,710 on 72 GPUs, Nebius 1,186,750 on 72 GPUs and NVIDIA 1,124,500 on 72 GPUs. 72 GPU entries top three: CoreWeave 1,197,710 on GB300, Nebius 1,186,750 on GB300 and NVIDIA 1,124,500 on GB300",CONSISTENT,"Read as the single-rack claim if a larger NVIDIA system exists. The first reading counts every NVIDIA entry in the file, the second only 72 GPU entries.",false,CoreWeave: CoreWeave Leads Cloud Providers in MLPerf Inference v6.1 Performance with NVIDIA Blackwell Ultra,https://coreweave.com/blog/coreweave-leads-cloud-providers-in-mlperf-r-inference-v6-1-performance-with-nvidia-blackwell-ultra,2026-10-28,MLPerf Inference v6.1
+CV2,CoreWeave,CoreWeave v6.1,"In server scenario, CoreWeave sustained 1,196 queries per second, the highest server throughput among cloud providers, and 1,135 samples per second in offline scenario.",true,CoreWeave 72 GPU GB300 Qwen3-VL entries against the other cloud provider entries,"1,196 Server (highest among cloud providers), 1,135 Offline","Server 1,195.59 (top cloud provider entry CoreWeave 1,195.59, next Oracle 1,146.39), Offline 1,135.25",CONSISTENT,"Our cloud provider list: Azure, CoreWeave, Crusoe, Google, Lambda, Nebius, Oracle. The post names no list.",false,CoreWeave: CoreWeave Leads Cloud Providers in MLPerf Inference v6.1 Performance with NVIDIA Blackwell Ultra,https://coreweave.com/blog/coreweave-leads-cloud-providers-in-mlperf-r-inference-v6-1-performance-with-nvidia-blackwell-ultra,2026-10-28,MLPerf Inference v6.1
+CV3,CoreWeave,CoreWeave v6.1,"comparing the 72-GPU v6.1 submission with the 64-GPU v6.0 submission on NVIDIA GB200 NVL72,derived per-GPU Server throughput increased by 19.8%",true,"CoreWeave GB200 NVL72 DeepSeek-R1 Server, v6.1 72 GPU entry over v6.0 64 GPU entry, each divided by its accelerator count",+19.8% per-GPU Server,"5,830.2 over 4,867.7, +19.77%",CONSISTENT,"Per-GPU is total divided by accelerators, a derived value the post marks as not verified by MLCommons.",false,CoreWeave: CoreWeave Leads Cloud Providers in MLPerf Inference v6.1 Performance with NVIDIA Blackwell Ultra,https://coreweave.com/blog/coreweave-leads-cloud-providers-in-mlperf-r-inference-v6-1-performance-with-nvidia-blackwell-ultra,2026-10-28,MLPerf Inference v6.1
+CV4,CoreWeave,CoreWeave v6.1,"CoreWeave sustained 16,635 tokens per second per GPU in offline scenario, and 16,118 in server scenario.",true,CoreWeave 72 GPU GB300 GPT-OSS-120B entries divided by 72,"16,635 Offline and 16,118 Server per GPU","16,634.9 and 16,117.9",CONSISTENT,,false,CoreWeave: CoreWeave Leads Cloud Providers in MLPerf Inference v6.1 Performance with NVIDIA Blackwell Ultra,https://coreweave.com/blog/coreweave-leads-cloud-providers-in-mlperf-r-inference-v6-1-performance-with-nvidia-blackwell-ultra,2026-10-28,MLPerf Inference v6.1
+CV5,CoreWeave,CoreWeave v6.1,"These were the highest per-GPU throughput of any v6.1 Datacenter Closed submission on GPT-OSS-120B, on any silicon, in both scenarios.",true,"closed v6.1 GPT-OSS-120B entries, result divided by accelerator count, all availability categories and available only",highest per-GPU in Offline and Server,"all categories: Offline top CoreWeave 16,635 then Nebius 16,530, Server top CoreWeave 16,118 then NVIDIA 16,022. Available only: Offline top CoreWeave, Server top CoreWeave",CONSISTENT,Per-GPU is a derived value. The post marks it not verified by MLCommons.,false,CoreWeave: CoreWeave Leads Cloud Providers in MLPerf Inference v6.1 Performance with NVIDIA Blackwell Ultra,https://coreweave.com/blog/coreweave-leads-cloud-providers-in-mlperf-r-inference-v6-1-performance-with-nvidia-blackwell-ultra,2026-10-28,MLPerf Inference v6.1
+CV6,CoreWeave,CoreWeave v6.1,CoreWeave achieved over 1.16 million tokens per second in server and over 1.19 million tokens per second in offline scenarios on a single NVIDIA GB300 NVL72 rack.,true,CoreWeave 72 GPU GB300 GPT-OSS-120B entries,"over 1.16 million Server, over 1.19 million Offline","Server 1,160,490, Offline 1,197,710",CONSISTENT,,false,CoreWeave: CoreWeave Leads Cloud Providers in MLPerf Inference v6.1 Performance with NVIDIA Blackwell Ultra,https://coreweave.com/blog/coreweave-leads-cloud-providers-in-mlperf-r-inference-v6-1-performance-with-nvidia-blackwell-ultra,2026-10-28,MLPerf Inference v6.1
+CV7,CoreWeave,CoreWeave v6.1,"Our NVIDIA GB200 NVL72 posted the highest rack-scale GPT-OSS-120B total in its class, reaching 901,058 tokens per second in server and 911,566 tokens per second in offline scenarios.",true,closed available v6.1 GPT-OSS-120B entries on 72 GB200 accelerators,"highest total in class, 901,058 Server and 911,566 Offline","Server top CoreWeave 901,058 of 2, Offline top CoreWeave 911,566 of 2",CONSISTENT,Our rule: class means 72 GB200 accelerators.,false,CoreWeave: CoreWeave Leads Cloud Providers in MLPerf Inference v6.1 Performance with NVIDIA Blackwell Ultra,https://coreweave.com/blog/coreweave-leads-cloud-providers-in-mlperf-r-inference-v6-1-performance-with-nvidia-blackwell-ultra,2026-10-28,MLPerf Inference v6.1
+CV8,CoreWeave,CoreWeave v6.1,Our NVIDIA HGX B200 and HGX B300 systems led all HGX submissions from cloud providers in both scenarios.,true,closed available v6.1 GPT-OSS-120B entries on eight B200 or eight B300 from cloud providers,led B200 and B300 in Server and Offline,"B200 Server top CoreWeave 90,205 of 2, B200 Offline top CoreWeave 91,487 of 2, B300 Server top CoreWeave 111,247 of 3, B300 Offline top CoreWeave 112,840 of 3",CONSISTENT,"Our cloud provider list: Azure, CoreWeave, Crusoe, Google, Lambda, Nebius, Oracle. Cisco, Dell, HPE and other server makers are not counted.",false,CoreWeave: CoreWeave Leads Cloud Providers in MLPerf Inference v6.1 Performance with NVIDIA Blackwell Ultra,https://coreweave.com/blog/coreweave-leads-cloud-providers-in-mlperf-r-inference-v6-1-performance-with-nvidia-blackwell-ultra,2026-10-28,MLPerf Inference v6.1
+CV9,CoreWeave,CoreWeave v6.1,"We delivered the highest throughput among cloud providers with NVIDIA GB300 NVL72 with a throughput of 944,902 tokens per second in server scenario and 1,136,100 in offline scenario.",true,closed available v6.1 Llama 2 70B entries from cloud providers on 72 GB300,"highest, 944,902 Server and 1,136,100 Offline","Server top CoreWeave 944,902 of 1 (CoreWeave 944,902). Offline top CoreWeave 1,136,100 of 1",CONSISTENT,"Llama 2 70B is read as llama2-70b-99.9, the only Llama 2 70B variant CoreWeave submitted. CoreWeave is the only cloud provider entry on 72 GB300 for this model in the file, so the ranking is by the absence of other entries.",false,CoreWeave: CoreWeave Leads Cloud Providers in MLPerf Inference v6.1 Performance with NVIDIA Blackwell Ultra,https://coreweave.com/blog/coreweave-leads-cloud-providers-in-mlperf-r-inference-v6-1-performance-with-nvidia-blackwell-ultra,2026-10-28,MLPerf Inference v6.1
+IP2,Intel,Intel v6.1,Customer and partner participation on Intel platforms also increased from 29 results in v6.0 to 39 in v6.1,true,"closed datacenter rows not submitted by Intel with a Xeon host and no accelerator, or an Arc Pro accelerator, in the v6.0 and v6.1 files","29 results in v6.0, 39 in v6.1","v6.0 20 rows (5 entries), v6.1 39 rows (11 entries)",BELOW,"Our rule follows the post footnote: Xeon as primary compute or Arc Pro as accelerator, Intel own submissions excluded. A result is one model and scenario row. The v6.1 count of 39 reproduces. In v6.0 five partners (Cisco, Dell, Lenovo, Quanta Cloud Technology, Supermicro) have four rows each, which is 20.",false,Intel: Intel Software Optimizations Boost AI Inference in MLPerf v6.1,https://www.intel.com/content/www/us/en/newsroom/news/data-center/intel-software-optimizations-boost-ai-inference-in-mlperf-v6-1.html,2026-10-28,MLPerf Inference v6.1
+IP4,Intel,Intel v6.1,"Intel broadened its Xeon participation in MLPerf v6.1 from two benchmarked Xeon 6 SKUs in v6.0 to five, increasing the number of CPU inference results from 24 to 35",true,"closed datacenter rows with no accelerator, v6.0 and v6.1, by processor model","two SKUs to five, 24 to 35 results","v6.0: 24 rows, 2 processor models. v6.1: 33 rows, 5 processor models (Intel(R) Xeon(R) 6787P, Intel(R) Xeon(R) 6788P, Intel(R) Xeon(R) 6972P, Intel(R) Xeon(R) 6980P, Intel(R) Xeon(R) 6987P-C)",BELOW,"Counting all submitters. The two SKU and five SKU counts reproduce for Intel own entries. The v6.1 file holds 33 CPU rows, not 35.",false,Intel: Intel Software Optimizations Boost AI Inference in MLPerf v6.1,https://www.intel.com/content/www/us/en/newsroom/news/data-center/intel-software-optimizations-boost-ai-inference-in-mlperf-v6-1.html,2026-10-28,MLPerf Inference v6.1
+IP1,Intel,Intel v6.1,Intel Xeon 6980P processors delivered a 2.4x increase in Llama 3.1 8B Server throughput and a 56% increase in Offline throughput,true,"Intel 1-node-2S-GNR_128C, v6.0 entry 6.0-0057 against the v6.1 entry, Llama 3.1 8B","2.4x Server (1,139.51 vs 471.54), +56% Offline (1,916.74 vs 1,229.56)","Server 1,139.51 against 471.54 (2.42x), Offline 1,916.74 against 1,229.56 (+55.9%)",CONSISTENT,"The post says increase of 2.4x. The ratio is 2.42x, which is +142%.",false,Intel: Intel Software Optimizations Boost AI Inference in MLPerf v6.1,https://www.intel.com/content/www/us/en/newsroom/news/data-center/intel-software-optimizations-boost-ai-inference-in-mlperf-v6-1.html,2026-10-28,MLPerf Inference v6.1
+IP3,Intel,Intel v6.1,"Oracle contributed its first Intel-based submission; Red Hat delivered its first Xeon CPU inference submission, and Quanta Cloud Technology and Supermicro provided the first partner submissions using Intel Arc Pro B70",true,"Oracle and Red Hat Xeon CPU-only entries and non-Intel Arc Pro B70 entries in v5.0, v5.1, v6.0 and v6.1",three firsts,"Oracle Xeon CPU-only rows: v6.1 3, earlier 0. RedHat_Intel rows: v6.1 3, earlier 0. Non-Intel B70 submitters in v6.1: Quanta_Cloud_Technology and Supermicro, in v6.0: 0",CONSISTENT,Intel-based for Oracle is read as a Xeon CPU-only entry. Oracle GPU systems with Intel hosts are not counted.,false,Intel: Intel Software Optimizations Boost AI Inference in MLPerf v6.1,https://www.intel.com/content/www/us/en/newsroom/news/data-center/intel-software-optimizations-boost-ai-inference-in-mlperf-v6-1.html,2026-10-28,MLPerf Inference v6.1
+IP5,Intel,Intel v6.1,Intel Xeon remains the only standalone server CPU represented in MLPerf Inference submissions,true,"closed datacenter rows with no accelerator, v6.0 and v6.1, host processor vendor",Xeon is the only standalone server CPU,processor vendors in CPU-only rows: Intel,CONSISTENT,Read for the v6.0 and v6.1 files only.,false,Intel: Intel Software Optimizations Boost AI Inference in MLPerf v6.1,https://www.intel.com/content/www/us/en/newsroom/news/data-center/intel-software-optimizations-boost-ai-inference-in-mlperf-v6-1.html,2026-10-28,MLPerf Inference v6.1
+IP6,Intel,Intel v6.1,"A single node configured with four Intel Arc Pro B70 GPUs provides 128GB of VRAM and supported submissions across Llama 3.1 8B, Llama 2 70B, gpt-oss-120B, Whisper and end-to-end retrieval-augmented generation (E2E-RAG)",true,Intel four B70 entries in the v6.1 file. 4 times 32GB per the post footnote,"four GPUs, 128GB, five model families","4 accelerators. Models e2e-rag-db, e2e-rag-qna, gpt-oss-120b, llama2-70b-99.9, llama3_1-8b and whisper. 4 x 32 = 128 GB",CONSISTENT,"The file splits E2E-RAG into two model names, database and question answering. The 32GB per card is from the post footnote, not a file field.",false,Intel: Intel Software Optimizations Boost AI Inference in MLPerf v6.1,https://www.intel.com/content/www/us/en/newsroom/news/data-center/intel-software-optimizations-boost-ai-inference-in-mlperf-v6-1.html,2026-10-28,MLPerf Inference v6.1
+IP7,Intel,Intel v6.1,"On the same four-GPU system used in v6.0, gpt-oss-120B Server performance improved 36%, while Offline performance increased 27%",true,"Intel 4 x Arc Pro B70 entry, v6.0 entry 6.0-0101 against the v6.1 entry, gpt-oss-120B","+36% Server (1,296.87 vs 951.67), +27% Offline (1,956.40 vs 1,536.90)","Server 951.67 to 1,296.87 (+36.3%). Offline 1,536.90 to 1,956.40 (+27.3%)",CONSISTENT,,false,Intel: Intel Software Optimizations Boost AI Inference in MLPerf v6.1,https://www.intel.com/content/www/us/en/newsroom/news/data-center/intel-software-optimizations-boost-ai-inference-in-mlperf-v6-1.html,2026-10-28,MLPerf Inference v6.1
+IP8,Intel,Intel v6.1,In a single measured run on a system combining an Intel Xeon 6787P processor with four Intel Arc Pro B70 GPUs,true,v6.1 E2E-RAG entries,Xeon 6787P with four Arc Pro B70,"2 E2E-RAG rows: Intel 1-node-2S-Xeon_6787P Intel(R) Xeon(R) 6787P 4 accelerators e2e-rag-db, Intel 1-node-2S-Xeon_6787P Intel(R) Xeon(R) 6787P 4 accelerators e2e-rag-qna",CONSISTENT,,false,Intel: Intel Software Optimizations Boost AI Inference in MLPerf v6.1,https://www.intel.com/content/www/us/en/newsroom/news/data-center/intel-software-optimizations-boost-ai-inference-in-mlperf-v6-1.html,2026-10-28,MLPerf Inference v6.1
+IP9,Intel,Intel v6.1,"The configuration lines in footnotes 1 and 3 of the post give 2x Intel Xeon 6980P processors with 128 cores and no discrete accelerator, and 1x Intel Xeon 698X processor with 86 cores and four Intel Arc Pro B70 GPUs.",false,"Intel entries in the v6.1 file, host processor fields","2x Xeon 6980P with 128 cores each, no accelerator. 1x Xeon 698X with 86 cores and 4x Arc Pro B70","2x Intel(R) Xeon(R) 6980P, 128 cores, 0 accelerators. 1x Intel(R) Xeon(R) 698X, 86 cores, 4 accelerators",CONSISTENT,,false,Intel: Intel Software Optimizations Boost AI Inference in MLPerf v6.1,https://www.intel.com/content/www/us/en/newsroom/news/data-center/intel-software-optimizations-boost-ai-inference-in-mlperf-v6-1.html,2026-10-28,MLPerf Inference v6.1
+AZ6,Azure,Azure v6.1,maintaining the highest Interactive performance among all reported submissions,true,"closed v6.1 DeepSeek-R1 Interactive entries, all availability categories",highest Interactive among all submissions,"all categories: top NVIDIA 652,750 (preview, NVIDIA Vera Rubin), then Azure 260,098. Available only: top Azure",BELOW,The higher entry is a preview category submission. Among available entries Azure is first.,false,"Azure: Azure Delivers Strong MLPerf Inference v6.1 Results with NVIDIA GB300 & GB200 (post dated 16 September 2026, read from a republished copy at argonsys.com)",https://argonsys.com/microsoft-cloud/library/azure-delivers-strong-mlperf-inference-v6-1-results-with-nvidia-gb300-gb200/,2026-10-28,MLPerf Inference v6.1
+AZ1,Azure,Azure v6.1,"Azure submitted GB200 results at 18-node (1-rack) and 72-node (4-rack) scale, and GB300 results at 18 nodes (1 rack)",true,"Azure entries in the v6.1 file, nodes and accelerators","GB200 at 18 and 72 nodes, GB300 at 18 nodes",GB200 72 accelerators on 18 nodes and 288 accelerators on 72 nodes. GB300 72 accelerators on 18 nodes,CONSISTENT,,false,"Azure: Azure Delivers Strong MLPerf Inference v6.1 Results with NVIDIA GB300 & GB200 (post dated 16 September 2026, read from a republished copy at argonsys.com)",https://argonsys.com/microsoft-cloud/library/azure-delivers-strong-mlperf-inference-v6-1-results-with-nvidia-gb300-gb200/,2026-10-28,MLPerf Inference v6.1
+AZ2,Azure,Azure v6.1,"The results table in the post lists seven DeepSeek-R1 throughput values in tokens per second: GB300 NVL72 on 18 nodes 260,098 Interactive, 661,143 Offline and 589,391 Server, GB200 NVL72 on 18 nodes 542,075 Offline and 426,796 Server, and GB200 NVL72 on 72 nodes 2,091,190 Offline and 1,598,510 Server.",false,Azure DeepSeek-R1 entries in the v6.1 file,seven table values in tokens/s,"GB300/72 Interactive 260,098, GB300/72 Offline 661,143, GB300/72 Server 589,391, GB200/72 Offline 542,075, GB200/72 Server 426,796, GB200/288 Offline 2,091,190, GB200/288 Server 1,598,510",CONSISTENT,Every table value matches the file to the digit printed.,false,"Azure: Azure Delivers Strong MLPerf Inference v6.1 Results with NVIDIA GB300 & GB200 (post dated 16 September 2026, read from a republished copy at argonsys.com)",https://argonsys.com/microsoft-cloud/library/azure-delivers-strong-mlperf-inference-v6-1-results-with-nvidia-gb300-gb200/,2026-10-28,MLPerf Inference v6.1
+AZ3,Azure,Azure v6.1,Azure achieved the highest reported Interactive score on GB300 NVL72,true,closed available v6.1 DeepSeek-R1 Interactive entries on GB300,highest Interactive on GB300,"top Azure 260,098, next NVIDIA 253,506",CONSISTENT,,false,"Azure: Azure Delivers Strong MLPerf Inference v6.1 Results with NVIDIA GB300 & GB200 (post dated 16 September 2026, read from a republished copy at argonsys.com)",https://argonsys.com/microsoft-cloud/library/azure-delivers-strong-mlperf-inference-v6-1-results-with-nvidia-gb300-gb200/,2026-10-28,MLPerf Inference v6.1
+AZ4,Azure,Azure v6.1,the highest reported Server and Offline scores on GB200 NVL72 at both 1- and 4-rack scale,true,closed available v6.1 DeepSeek-R1 entries on 72 and 288 GB200 accelerators,highest Server and Offline at one rack and four racks,"one rack Server top Azure of 4, Offline top Azure of 4. Four racks Server 1 entry (Azure), Offline 1 entry",CONSISTENT,"Azure is the only entry at four racks, so that part holds by the absence of others.",false,"Azure: Azure Delivers Strong MLPerf Inference v6.1 Results with NVIDIA GB300 & GB200 (post dated 16 September 2026, read from a republished copy at argonsys.com)",https://argonsys.com/microsoft-cloud/library/azure-delivers-strong-mlperf-inference-v6-1-results-with-nvidia-gb300-gb200/,2026-10-28,MLPerf Inference v6.1
+AZ5,Azure,Azure v6.1,Azure's GB300 submission achieved 97.7% of the best reported Server performance and 95.8% of the best reported Offline performance,true,"Azure GB300 entry against the best 72 GPU GB300 DeepSeek-R1 entry, and against the best entry of any kind","97.7% of best Server, 95.8% of best Offline","against the best 72 GPU GB300 entry (Nebius 603,023 and Nebius 689,961): 97.7% and 95.8%. Against the best entry of any kind (Crusoe 2,405,310 Server, Crusoe 2,901,950 Offline): 24.5% and 22.8%",CONSISTENT,Our rule: best reported means the best 72 GPU GB300 entry. Against Azure's own 288 GPU GB200 entry the shares are about 37% and 32%.,false,"Azure: Azure Delivers Strong MLPerf Inference v6.1 Results with NVIDIA GB300 & GB200 (post dated 16 September 2026, read from a republished copy at argonsys.com)",https://argonsys.com/microsoft-cloud/library/azure-delivers-strong-mlperf-inference-v6-1-results-with-nvidia-gb300-gb200/,2026-10-28,MLPerf Inference v6.1
+AZ7,Azure,Azure v6.1,marking our first rack-scale and cross-rack MLPerf Inference submissions,true,"Azure entries in the v5.0, v5.1 and v6.0 files, accelerator counts",first rack-scale submission,earlier Azure entries: v5.1 entry 5.1-0008 with 4 accelerators. System files in v3.1 and v4.0: v3.1 ND_H100_v5_TRT.json: 1 node of 8 accelerators and v4.0 NC_H100_v5_TRT.json: 1 node of 2 accelerators. 72 or more accelerators before v6.1: 0,CONSISTENT,"Rounds before v5.0 were read from system description files, which show one node of 2 to 8 accelerators.",false,"Azure: Azure Delivers Strong MLPerf Inference v6.1 Results with NVIDIA GB300 & GB200 (post dated 16 September 2026, read from a republished copy at argonsys.com)",https://argonsys.com/microsoft-cloud/library/azure-delivers-strong-mlperf-inference-v6-1-results-with-nvidia-gb300-gb200/,2026-10-28,MLPerf Inference v6.1
+CX1,Cisco,Cisco X210c v5.1,v5.1 Data Center Inference results obtained on a Cisco UCS X210c M8 Compute Node powered by Intel Xeon 6 processor family,true,v5.1 entry 5.1-0017,v5.1 datacenter results on a UCS X210c M8 with Xeon 6,"Cisco X210M8-1-node-2S-GNR_86C, closed division, INTEL(R) XEON(R) 6787P, 8 rows across 5 models",CONSISTENT,The system is named X210M8-1-node-2S-GNR_86C in the file.,false,Cisco: AI Performance: MLPerf Inference on Cisco UCS X210c M8 Compute Node with Intel Xeon 6 Processor (white paper updated 2 October 2025),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-x210c-m8-compute-node-wp.html,2026-10-28,MLPerf Inference v5.1
+CX2,Cisco,Cisco X210c v5.1,"Table 6 in the post lists inferences per second for five benchmarks: DLRM-v2-99.9 12503.3 Offline and 11801.67 Server, Retinanet 501.263 and 400.42, RGAT 16102.2 Offline, Whisper 1418.6 Offline, and Llama 3.1 8B 819.624 and 257.75.",false,entry 5.1-0017 result field,"eight values, DLRM-v2-99.9 12503.3 and 11801.67, Retinanet 501.263 and 400.42, RGAT 16102.2, Whisper 1418.6, Llama 3.1 8B 819.624 and 257.75","dlrm-v2-99.9 Offline 12503.3, dlrm-v2-99.9 Server 11801.7, retinanet Offline 501.263, retinanet Server 400.419, rgat Offline 16102.2, whisper Offline 1418.6, llama3.1-8b Offline 819.624, llama3.1-8b Server 257.749",CONSISTENT,"The file prints DLRM Server as 11801.7, Retinanet Server as 400.419 and Llama 3.1 8B Server as 257.749. The post prints the same values rounded differently. Our rule: within 0.06 above 1,000 and within 0.006 below.",false,Cisco: AI Performance: MLPerf Inference on Cisco UCS X210c M8 Compute Node with Intel Xeon 6 Processor (white paper updated 2 October 2025),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-x210c-m8-compute-node-wp.html,2026-10-28,MLPerf Inference v5.1
+CX3,Cisco,Cisco X210c v5.1,Processor name Intel Xeon 6787P Host processors per node 2 Host processor core count 86,true,entry 5.1-0017 host fields,"2 x Xeon 6787P, 86 cores","2 x INTEL(R) XEON(R) 6787P, 86 cores",CONSISTENT,,false,Cisco: AI Performance: MLPerf Inference on Cisco UCS X210c M8 Compute Node with Intel Xeon 6 Processor (white paper updated 2 October 2025),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-x210c-m8-compute-node-wp.html,2026-10-28,MLPerf Inference v5.1
+CC10,Cisco,Cisco C845A v6.1,CPU 2x AMD EPYC 9575 64-Core Processor Number of cores 64,true,"Cisco C845A entries in the v5.1 and v6.1 files, host processor fields","2 x AMD EPYC 9575, 64 cores","v5.1 2 x AMD EPYC 9575F 64-Core Processor, 64 cores. v6.1 2 x AMD EPYC 9655, 96 cores",BELOW,The v5.1 entries list EPYC 9575F with 64 cores. The v6.1 entries list EPYC 9655 with 96 cores. Table 2 in the post gives one CPU for both rounds.,false,Cisco: AI Performance: MLPerf Inference on Cisco UCS C845A M8 Rack Server with NVIDIA GPUs (white paper updated 5 October 2026),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-c845a-m8-rack-server-wp.html,2026-10-28,MLPerf Inference v5.1 and v6.1
+CC2,Cisco,Cisco C845A v6.1,Cisco has tested most of the datasets listed in Table 1 on the Cisco UCS C845A M8 Rack Server and submitted the results to MLCommons,true,C845A M8 models in the v5.1 file against the seven models in Table 1 of the post,most of seven models submitted,2 of 7 models in v5.1: llama3.1-8b and retinanet,BELOW,"The post then marks the Llama 2 70B, Stable Diffusion XL and Whisper results as collected after the deadline and not verified. The file holds only Llama 3.1 8B and Retinanet for this server.",false,Cisco: AI Performance: MLPerf Inference on Cisco UCS C845A M8 Rack Server with NVIDIA GPUs (white paper updated 5 October 2026),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-c845a-m8-rack-server-wp.html,2026-10-28,MLPerf Inference v5.1 and v6.1
+CC4,Cisco,Cisco C845A v6.1,Figure 16 shows the performance of the gpt-oss-120b model tested on a Cisco UCS C880A M8 Rack Server with 8x NVIDIA RTX PRO 6000 GPUs.,true,"Cisco gpt-oss-120b entries with eight RTX PRO 6000 in the v6.1 file, system name",C880A M8 with 8 RTX PRO 6000,"Cisco UCS C845A M8 (8x RTXPro6000-PCIe-96GB, TensorRT). The C880A M8 entries in the file use B300",BELOW,The figure title in the post says C845A and the sentence says C880A. The file names the eight RTX PRO 6000 entry C845A M8.,false,Cisco: AI Performance: MLPerf Inference on Cisco UCS C845A M8 Rack Server with NVIDIA GPUs (white paper updated 5 October 2026),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-c845a-m8-rack-server-wp.html,2026-10-28,MLPerf Inference v5.1 and v6.1
+CC6,Cisco,Cisco C845A v6.1,"The Cisco UCS C845A M8 platform with 8x NVIDIA H200 NVL GPUs emerged as the leader, securing first position for the Retinanet model.",true,Cisco 5.1-0011 against closed datacenter v5.1 entries with eight H200 NVL,first for Retinanet,"Offline 3/3, Server 1/3",BELOW,"Our rank rule: entries with the same accelerator model and count, higher result first. Overall among all closed entries for the model and scenario the rank is lower. First holds in Server and not in Offline.",false,Cisco: AI Performance: MLPerf Inference on Cisco UCS C845A M8 Rack Server with NVIDIA GPUs (white paper updated 5 October 2026),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-c845a-m8-rack-server-wp.html,2026-10-28,MLPerf Inference v5.1 and v6.1
+CC1,Cisco,Cisco C845A v6.1,"Inference 5.1 and 6.1: Datacenter performance was measured on a Cisco UCS C845A M8 Rack Server equipped with 8 NVIDIA H200 NVL GPUs, 8 NVIDIA L40S PCIe GPUs and 8 NVIDIA RTX PRO 6000 and RTX PRO 4500 PCIe GPUs.",true,Cisco C845A M8 entries in the v5.1 and v6.1 files,four eight-GPU configurations,"v5.1: H200 NVL 4 rows, L40S 2 rows. v6.1: RTX PRO 6000 8 rows, RTX PRO 4500 4 rows",CONSISTENT,,false,Cisco: AI Performance: MLPerf Inference on Cisco UCS C845A M8 Rack Server with NVIDIA GPUs (white paper updated 5 October 2026),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-c845a-m8-rack-server-wp.html,2026-10-28,MLPerf Inference v5.1 and v6.1
+CC3,Cisco,Cisco C845A v6.1,"Note: For Llama2-70b performance data, the results have not been verified by MLCommons Association because the results were collected after the MLPerf submission deadline. ... Note: For NVIDIA RTX PRO 4500 Llama2-70b performance data, the results have not been verified by MLCommons Association because the results were collected after the MLPerf submission deadline.",true,"C845A M8 entries in the v5.1 and v6.1 files for Llama 2 70B, Stable Diffusion XL and Whisper on H200 NVL, and Llama 2 70B on RTX PRO 4500",these results are not published entries,"v5.1 C845A rows: Llama 2 70B 0, Whisper 0, Stable Diffusion XL 0. v6.1 RTX PRO 4500 Llama 2 70B rows: 0",CONSISTENT,"None of the unverified results are entries in the file, which is what the notes say.",false,Cisco: AI Performance: MLPerf Inference on Cisco UCS C845A M8 Rack Server with NVIDIA GPUs (white paper updated 5 October 2026),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-c845a-m8-rack-server-wp.html,2026-10-28,MLPerf Inference v5.1 and v6.1
+CC5,Cisco,Cisco C845A v6.1,"The Cisco UCS C845A M8 platform with 8x NVIDIA H200 NVL GPUs emerged as the leader, securing first position for the Llama3.1-8b model.",true,Cisco 5.1-0011 against closed datacenter v5.1 entries with eight H200 NVL,first for Llama 3.1 8B,"Offline 1/1, Server 1/1. The only eight H200 NVL entry for this model",CONSISTENT,"Our rank rule: entries with the same accelerator model and count, higher result first. Overall among all closed entries for the model and scenario the rank is lower. Cisco is the only entry in this group, so first holds by the absence of others.",false,Cisco: AI Performance: MLPerf Inference on Cisco UCS C845A M8 Rack Server with NVIDIA GPUs (white paper updated 5 October 2026),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-c845a-m8-rack-server-wp.html,2026-10-28,MLPerf Inference v5.1 and v6.1
+CC7,Cisco,Cisco C845A v6.1,"The Cisco UCS C845A M8 platform with 8x NVIDIA RTX PRO 6000 GPUs emerged as the leader, securing first position for the Llama2-70b model for Server scenario",true,Cisco C845A RTX PRO 6000 llama2-70b-99.9 against closed available v6.1 entries with eight RTX PRO 6000,first for Llama 2 70B Server,"Server 1/4, Offline 2/4",CONSISTENT,"Our rank rule: entries with the same accelerator model and count, higher result first. Overall among all closed entries for the model and scenario the rank is lower. The file has the 99.9 variant for this server. Offline is second.",false,Cisco: AI Performance: MLPerf Inference on Cisco UCS C845A M8 Rack Server with NVIDIA GPUs (white paper updated 5 October 2026),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-c845a-m8-rack-server-wp.html,2026-10-28,MLPerf Inference v5.1 and v6.1
+CC8,Cisco,Cisco C845A v6.1,"The Cisco UCS C845A M8 platform with 8x NVIDIA RTX PRO 4500 GPUs emerged as the leader, securing first position for the Llama3.1-8b model.",true,Cisco C845A RTX PRO 4500 llama3.1-8b against closed available v6.1 entries with eight RTX PRO 4500,first for Llama 3.1 8B,"Interactive 1/1, Offline 1/1, Server 1/1",CONSISTENT,"Our rank rule: entries with the same accelerator model and count, higher result first. Overall among all closed entries for the model and scenario the rank is lower. Cisco is the only entry in this group, so first holds by the absence of others.",false,Cisco: AI Performance: MLPerf Inference on Cisco UCS C845A M8 Rack Server with NVIDIA GPUs (white paper updated 5 October 2026),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-c845a-m8-rack-server-wp.html,2026-10-28,MLPerf Inference v5.1 and v6.1
+CC9,Cisco,Cisco C845A v6.1,"The Cisco UCS C845A M8 platform with 8x NVIDIA RTX PRO 4500 GPUs emerged as the leader, securing first position for the Whisper model.",true,Cisco C845A RTX PRO 4500 whisper against closed available v6.1 entries with eight RTX PRO 4500,first for Whisper,Offline 1/1,CONSISTENT,"Our rank rule: entries with the same accelerator model and count, higher result first. Overall among all closed entries for the model and scenario the rank is lower. Cisco is the only entry in this group, so first holds by the absence of others.",false,Cisco: AI Performance: MLPerf Inference on Cisco UCS C845A M8 Rack Server with NVIDIA GPUs (white paper updated 5 October 2026),https://www.cisco.com/c/en/us/products/collateral/servers-unified-computing/ucs-c-series-rack-servers/mlperf-ucs-c845a-m8-rack-server-wp.html,2026-10-28,MLPerf Inference v5.1 and v6.1
+HV13,HPE,HPE v5.1,These results are 12% higher (Server) and 16% (Offline) than the nearest competitor's entry using the same number of NVIDIA H200 SXM GPUs.,true,HPE entry 5.1-0049 Llama3.1-8B against the next entry with eight H200 SXM,"+12% Server, +16% Offline","Server 64,914.6 over 57,970.4 (Quanta_Cloud_Technology), +12.0%. Offline 66,036.8 over 58,304.3 (Cisco), +13.3%",BELOW,"Server reproduces at 12.0%. Offline is 13.3%, not 16%.",false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV2,HPE,HPE v5.1,the DL380a achieved two overall #1 spots in the DLRM-v2-99 and DLRM-v2-99.9 benchmarks (Server scenario),true,"HPE entry 5.1-0050 ranked among closed datacenter v5.1 DLRM Server entries, by result per accelerator and by total result",overall #1 in DLRM-v2-99 and DLRM-v2-99.9 Server,per accelerator: rank 4 of 7 and 3 of 13. By total result: rank 1 of 7 and 1 of 13,BELOW,The post quotes the figures per GPU. By total system result HPE is first in DLRM-v2-99 Server. Per GPU it is not first in either.,false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV7,HPE,HPE v5.1,6 MLPerf Inference: Datacenter v5.0 Closed. Llama3.1-8b benchmark. Submission ID 5.0-0051 ... 7 MLPerf Inference: Datacenter v5.0 Closed. Llama3.1-8b benchmark. Submission ID 5.0-0011 ... 8 MLPerf Inference: Datacenter v5.0 Closed. Llama2 70B benchmark. Submission ID 5.0-0038,true,"entries 5.0-0051, 5.0-0011 and 5.0-0038 in the v5.0 file","footnotes 6 and 7 cite v5.0 Llama3.1-8b entries, footnote 8 cites an HPE DL380a Llama 2 70B entry","5.0-0051 is Lenovo ThinkSystem SR680a V3(8x H200-SXM-141GB, TensorRT). 5.0-0011 is Cisco Cisco UCS C885A M8 (8x H200-SXM-141GB, TensorRT). v5.0 Llama3.1-8b rows in the file: 0. 5.0-0038 is HPE HPE ProLiant Compute DL384 Gen12 (2x GH200-144GB_aarch64, TensorRT)",BELOW,"The v5.0 file has no Llama3.1-8b rows. The 46,060.0 and 38,696.9 results are v5.1 entries 5.1-0051 and 5.1-0011. The DL380a L40S Llama 2 70B entry is 5.0-0046.",false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV1,HPE,HPE v5.1,"the DL380a achieved two overall #1 spots in the DLRM-v2-99 and DLRM-v2-99.9 benchmarks (Server scenario) with 65,021 and 41,357 queries/second per GPU, respectively",true,"HPE entry 5.1-0050 (10 H200 NVL), DLRM Server result divided by 10, then ranked among closed datacenter v5.1 entries","#1 in both, 65,021 and 41,357 per GPU","65,021 and 41,356.5 per GPU",CONSISTENT,Values are the file results divided by ten accelerators.,false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV10,HPE,HPE v5.1,"The HPE ProLiant DL384 Gen12, equipped with NVIDIA GH200 NVL2 accelerators, stood out in the DLRM-v2-99 benchmark, achieving 161,030 queries per second (Server scenario) and 174,456 samples per second (Offline scenario), making HPE the only vendor to submit results with this advanced configuration.",true,HPE entry 5.1-0053 and every v5.1 entry with a GH200 accelerator,"161,030 Server, 174,456 Offline, only vendor","161,030 and 174,456. v5.1 submitters with GH200: HPE",CONSISTENT,,false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV11,HPE,HPE v5.1,"In the edge scenario, HPE ProLiant ML30 Gen11 displayed strong throughput performance in RetinaNet benchmark tests, delivering 258.352 samples/second and multistream latency = 29.96 ms on a single NVIDIA RTX 4000 Ada 20GB GPU.",true,"HPE entry 5.1-0054, Retinanet Offline and MultiStream","258.352 samples/s, 29.96 ms, one RTX 4000 Ada","258.352 Offline, 29.96 ms MultiStream, 1 accelerator, NVIDIA RTX 4000 Ada-PCIe-20GB",CONSISTENT,,false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV12,HPE,HPE v5.1,"#1 in RetinaNet (Offline) with 14,996.6 samples per second ... # 1 in the new Llama3.1-8B (Server and Offline), with 64,914.6 and 66,036.8 tokens per second ... #1 in Mixtral-8x7B (Server and Offline), delivering 60,955.1 and 62,108.4 tokens per second ... #1 in the new Whisper (Offline) with 34,450.7 samples per second",true,"HPE entry 5.1-0049 (8 H200 SXM) against closed datacenter v5.1 entries with eight H200 SXM, six model and scenario pairs",six first places with the printed values,"retinanet Offline 14,996.6 (HPE, next 14,960.7), llama3.1-8b Server 64,914.6 (HPE, next 57,970.4), llama3.1-8b Offline 66,036.8 (HPE, next 58,304.3), mixtral-8x7b Server 60,955.1 (HPE, next 59,992.8), mixtral-8x7b Offline 62,108.4 (HPE, next 61,372.4), whisper Offline 34,450.7 (HPE, next 34,088.7)",CONSISTENT,Our rule: first among the entries with eight H200 SXM accelerators.,false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV14,HPE,HPE v5.1,"The only system featuring NVIDIA H200 GPUs to publish DeepSeek-R1 Offline performance of 8904.83 tokens per second, which was submitted in the Datacenter-Open category.",true,v5.1 DeepSeek-R1 Offline entries on H200,"only H200 DeepSeek-R1 Offline entry, 8904.83, open division","1 entries: HPE 5.1-0375 open 8,904.83",CONSISTENT,,false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV16,HPE,HPE v5.1,following a #1 result in this same model during the prior round of benchmark tests,true,"v5.0 Retinanet Offline entries with eight H100 SXM, footnote 16 IDs 5.0-0039 and 5.0-0040",#1 in v5.0 among eight H100 SXM,"top three: HPE 14,637.8 (5.0-0039), HPE 14,633.1 (5.0-0040) and Cisco 14,476.3 (5.0-0010)",CONSISTENT,"Footnote 16 names HPE entries 5.0-0039 and 5.0-0040, which are the top two.",false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV5,HPE,HPE v5.1,"In the MLPerf Inference v5.1 Llama3.1 8B test (Server scenario), the DL380a Gen12 claimed the top spot among systems with eight PCIe-based GPUs, delivering an impressive 46,060.0 tokens/second and outperforming Cisco's UCS C845A M8 by 19%, with its 38,696.9 tokens/second result.",true,"HPE entry 5.1-0051 and Cisco entry 5.1-0011, Llama 3.1 8B Server, against closed datacenter v5.1 entries with eight non-SXM accelerators","top among eight PCIe GPU systems, 46,060.0 against 38,696.9, +19%","HPE 46,060.0 (HPE ProLiant Compute DL380a Gen12 (8x RTX Pro 6000 Blackwell Server Edition)), Cisco 38,696.9, +19.0%. Top of 2 eight-GPU PCIe entries: HPE 46,060.0, next Cisco 38,696.9",CONSISTENT,Our rule: PCIe means an accelerator that is not SXM and not a Grace system.,false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV6,HPE,HPE v5.1,"Both the DL380a Gen12 and Dell's PowerEdge XE7745, equipped with eight NVIDIA L40S GPUs, competed in this benchmark, with the DL380a delivering 3655.89 tokens/second compared to Dell's 3481.53 tokens/second.",true,"HPE entry 5.0-0046 and Dell entry 5.0-0018, Llama 2 70B 99 Offline","3655.89 and 3481.53, only two eight-L40S entries","3,655.89 and 3,481.53. eight-L40S entries in v5.0: HPE 3,655.89 and Dell 3,481.53",CONSISTENT,,false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV8,HPE,HPE v5.1,"the HPE ProLiant DL385 Gen11 immediately secured a #1 spot, delivering the best per-GPU performance for a PCIe-based system in the new speech-to-text Whisper benchmark, with 3962.78 samples/second per GPU, when equipped with NVIDIA H200 NVL 141GB GPUs",true,"HPE entry 5.1-0052 Whisper Offline divided by 2, against closed datacenter v5.1 Whisper entries with non-SXM accelerators","best per-GPU for a PCIe system, 3962.78 per GPU","3,962.78 per GPU. Top PCIe entry per GPU HPE 3,962.78 (5.1-0052), next Dell 3,959.26",CONSISTENT,Our rule: PCIe means an accelerator that is not SXM and not a Grace system.,false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV9,HPE,HPE v5.1,HPE was the only company to submit a server utilizing the NVIDIA RTX PRO 6000 Blackwell Server Edition GPU in this round of MLPerf Inference testing.,true,v5.1 file entries naming RTX PRO 6000,only HPE,submitters: HPE,CONSISTENT,,false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV15,HPE,HPE v5.1,demonstrated leadership in AI inferencing with fourteen #1 results across multiple scenarios in the Datacenter and Edge categories,true,the first-place results reproduced in this script,fourteen #1 results,"the post counts eight ProLiant results and six XD670 results. Six XD670 results reproduce (HV12). The ProLiant ones are not all reproduced by our rule: per accelerator DLRM is not first (HV2), Llama3.1-8B Server and Whisper per GPU are (HV5, HV8)",CANNOT_TELL,"The rules used for the other first places, such as Xeon with NVIDIA groups in Figure 2, are in figures, not text.",false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV3,HPE,HPE v5.1,where it performed 57% better than the next-best submission in the DLRM-v2-99 Offline scenario,true,HPE entry 5.0-0043 DLRM-v2-99 Offline against the other closed datacenter v5.0 entries,57% better than the next-best,"512,363. Rank 9 of 22 by total, 14 of 22 per accelerator. The highest other entry is 654,489 (-21.7%)",CANNOT_TELL,"No ranking by total or per accelerator gives +57%. The group the post compares against is not named, so the file cannot settle it.",false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+HV4,HPE,HPE v5.1,"In this new round, v5.1, it continued to dominate, outperforming competitors by 29% in the DLRM-v2-99 Server test, showcasing consistent excellence over multiple benchmark iterations.",true,HPE entry 5.1-0050 DLRM-v2-99 Server against the next closed datacenter v5.1 entry,outperforming competitors by 29%,"650,211 against the next entry 591,162 (+10.0%)",CANNOT_TELL,The next entry by total result is 10% lower. The group the 29% is taken against is not named.,false,HPE: HPE delivers several world records in latest MLPerf Inference benchmarks (post dated 9 September 2025),https://community.hpe.com/t5/ai-unlocked/hpe-delivers-several-world-records-in-latest-mlperf-inference/ba-p/7255374,2026-10-28,MLPerf Inference v5.1
+SP1,Supermicro,Supermicro v5.0,Supermicro demonstrated more than 3 times the tokens per second (Token/s) generation for Llama2-70B and Llama3.1-405B benchmarks compared to H200 8-GPU systems.,true,"best Supermicro B200 entry against the best eight H200 SXM entry in the v5.0 file, four model and scenario rows",more than 3 times for Llama 2 70B and Llama 3.1 405B,"llama2-70b-99 Offline 98,782.4 over 35,453.0 = 2.79x, llama2-70b-99 Server 97,629.8 over 33,071.5 = 2.95x, llama3.1-405b Offline 1,538.2 over 573.6 = 2.68x, llama3.1-405b Server 1,080.3 over 294.3 = 3.67x",BELOW,Llama 2 70B is read as llama2-70b-99. Server for the 405B model is above 3x. The other three rows are 2.7x to 3.0x.,false,"Supermicro: Industry's First-to-Market Supermicro NVIDIA HGX B200 Systems Demonstrate AI Performance Leadership on MLPerf Inference v5.0 Results (press release, 3 April 2025)",https://www.supermicro.com/en/pressreleases/first-to-market-supermicro-nvidia-hgx-b200-systems-performance-leadership-mlperf-results,2026-10-28,MLPerf Inference v5.0
+SP3,Supermicro,Supermicro v5.0,Llama3.1-405b (offline) SYS-421GE-NBRT-LCC (8xB200-SXM-180GB) #1 Tokens/s 1521.74,true,"Supermicro entries 5.0-0072 and 5.0-0074, Llama 3.1 405B Offline, against closed datacenter v5.0 entries with eight B200",#1 at 1521.74 on the liquid-cooled system,"1521.74 is entry 5.0-0074 (SYS-A21GE-NBRT), not the liquid-cooled 5.0-0072, which is 1,538.17. Eight-B200 Offline order: 1,538.17 Supermicro 5.0-0072, 1,526.29 NVIDIA 5.0-0056, 1,521.74 Supermicro 5.0-0074, 1,498.56 Google 5.0-0078",BELOW,"1521.74 is third among eight-B200 entries behind 1538.17 (Supermicro, liquid-cooled) and 1526.29 (NVIDIA). The liquid-cooled system is first at 1538.17.",false,"Supermicro: Industry's First-to-Market Supermicro NVIDIA HGX B200 Systems Demonstrate AI Performance Leadership on MLPerf Inference v5.0 Results (press release, 3 April 2025)",https://www.supermicro.com/en/pressreleases/first-to-market-supermicro-nvidia-hgx-b200-systems-performance-leadership-mlperf-results,2026-10-28,MLPerf Inference v5.0
+SP2,Supermicro,Supermicro v5.0,"Stable Diffusion XL (Server) SYS-A21GE-NBRT (8x B200-SXM-180GB) #1 queries/s, 28.92 ... llama2-70b-interactive-99 (Server) SYS-A21GE-NBRT (8x B200-SXM-180GB) #1 Tokens/s, 62,265.70 ... Llama3.1-405b (Server) SYS-A21GE-NBRT (8x B200-SXNM-180GB) #1 Tokens/s, 1080.31 (for an 8-GPU node) ... mixtral-8x7b (Server) SYS-421GE-NBRT-LCC (8x B200-SXM-180GB) #1 Tokens/s, 129,047.00 ... mixtral-8x7b (Offline) SYS-421GE-NBRT-LCC (8x B200-SXM-180GB) #1 Tokens/s, 128,795.00",true,"Supermicro entries 5.0-0074 (SYS-A21GE-NBRT) and 5.0-0072 (SYS-421GE-NBRT-LCC), five rows, ranked among closed datacenter v5.0 entries with eight accelerators",five #1 results with the printed values and systems,"values 28.9204, 62265.7, 1080.31, 129047.0, 128795.0. First place entries 5.0-0074, 5.0-0074, 5.0-0074, 5.0-0072 and 5.0-0072",CONSISTENT,Our rule: first among the entries with eight accelerators. Each row matches the entry and system the post names.,false,"Supermicro: Industry's First-to-Market Supermicro NVIDIA HGX B200 Systems Demonstrate AI Performance Leadership on MLPerf Inference v5.0 Results (press release, 3 April 2025)",https://www.supermicro.com/en/pressreleases/first-to-market-supermicro-nvidia-hgx-b200-systems-performance-leadership-mlperf-results,2026-10-28,MLPerf Inference v5.0
+SP4,Supermicro,Supermicro v5.0,Supermicro is the only system vendor publishing record MLPerf inference performance (on select benchmarks) for both the air-cooled and liquid-cooled NVIDIA HGX B200 8-GPU systems.,true,submitters of eight B200 entries in the v5.0 file. Supermicro entries 5.0-0072 (LCC in the name) and 5.0-0074,only system vendor with both cooling types,"submitters with eight B200: Google, NVIDIA and Supermicro. Supermicro has two entries (5.0-0072, 5.0-0074)",CONSISTENT,Cooling type is read from the system name only. Google and NVIDIA are the other two submitters and are not system vendors in the post sense.,false,"Supermicro: Industry's First-to-Market Supermicro NVIDIA HGX B200 Systems Demonstrate AI Performance Leadership on MLPerf Inference v5.0 Results (press release, 3 April 2025)",https://www.supermicro.com/en/pressreleases/first-to-market-supermicro-nvidia-hgx-b200-systems-performance-leadership-mlperf-results,2026-10-28,MLPerf Inference v5.0
+SP6,Supermicro,Supermicro v5.0,"showed performance leadership running the Mixtral 8x7B Inference, Mixture of Experts benchmarks with 129,000 tokens/second ... delivered over 1,000 tokens/second inference for the large Llama3.1-405b model",true,"Supermicro B200 entries, Mixtral 8x7B and Llama 3.1 405B","129,000 tokens/s on Mixtral, over 1,000 tokens/s on Llama 3.1 405B","Mixtral Server 129,047, Offline 128,795. Llama 3.1 405B Server 1,080.31, Offline 1,538.17",CONSISTENT,,false,"Supermicro: Industry's First-to-Market Supermicro NVIDIA HGX B200 Systems Demonstrate AI Performance Leadership on MLPerf Inference v5.0 Results (press release, 3 April 2025)",https://www.supermicro.com/en/pressreleases/first-to-market-supermicro-nvidia-hgx-b200-systems-performance-leadership-mlperf-results,2026-10-28,MLPerf Inference v5.0
+SP7,Supermicro,Supermicro v5.0,"using the LLAMA2-70b benchmark, a Supermicro system with the NVIDIA HGX B200 SXM-180GB installed shows the highest performance from a Tier 1 system supplier",true,"closed datacenter v5.0 Llama 2 70B 99 Offline entries with eight B200, by submitter",highest Llama 2 70B performance from a Tier 1 system supplier,"eight-B200 Offline order: 98,858 NVIDIA, 98,782 Supermicro, 97,417 Supermicro, 97,124 Google. Best from a submitter other than Google and NVIDIA: Supermicro",CONSISTENT,"Tier 1 is not defined. Supermicro is the only B200 submitter that is a system vendor. NVIDIA is first at 98,858.",false,"Supermicro: Industry's First-to-Market Supermicro NVIDIA HGX B200 Systems Demonstrate AI Performance Leadership on MLPerf Inference v5.0 Results (press release, 3 April 2025)",https://www.supermicro.com/en/pressreleases/first-to-market-supermicro-nvidia-hgx-b200-systems-performance-leadership-mlperf-results,2026-10-28,MLPerf Inference v5.0
+SP5,Supermicro,Supermicro v5.0,"Within the operating margin, the Supermicro air-cooled NVIDIA HGX B200 system exhibited the same level of performance as the liquid-cooled HGX B200 system.",true,"air-cooled 5.0-0074 against liquid-cooled 5.0-0072, eight model and scenario rows",same performance within the operating margin,"llama2-70b-99 Offline -1.4%, llama2-70b-99 Server -0.1%, llama3.1-405b Offline -1.1%, llama3.1-405b Server +2.2%, mixtral-8x7b Offline -2.5%, mixtral-8x7b Server -0.1%, llama2-70b-interactive-99 Server +4.6%. Largest gap 4.6%",CANNOT_TELL,The operating margin is not defined in the post. Rows where both systems have a result are compared. The air-cooled system is higher in some rows and lower in others.,false,"Supermicro: Industry's First-to-Market Supermicro NVIDIA HGX B200 Systems Demonstrate AI Performance Leadership on MLPerf Inference v5.0 Results (press release, 3 April 2025)",https://www.supermicro.com/en/pressreleases/first-to-market-supermicro-nvidia-hgx-b200-systems-performance-leadership-mlperf-results,2026-10-28,MLPerf Inference v5.0
+AT1,Azure,Azure Training v6.0,"we scaled Llama 3.1 405B (NVFP4) pretraining to 8,192 GB200 GPUs across 128 racks on Azure's Fairwater infrastructure, converging in 7.07 minutes",true,"Azure llama31_405b entry in the v6.0 training times, accelerator count and official-style time","8,192 GB200 GPUs, 7.07 minutes","8192 GPUs (NVIDIA Blackwell GPU (GB200)), runs [7.212, 6.642, 7.069] minutes, official-style time 7.069. 64 GPUs per rack over 128 racks",CONSISTENT,"Racks and NVFP4 are not file fields. The file system name is 2048xAzure_GB200, which is 2,048 nodes of four GPUs.",false,"Azure: Inside Llama 3.1 405B MLPerf Training on Azure: System-Level Insights at 8K+ GPU Scale (post dated 24 June 2026, read from a republished copy on the Azure Feeds listing)",,2026-10-28,MLPerf Training v6.0
+AT2,Azure,Azure Training v6.0,"Azure's Llama 3.1 405B submission, which achieved a leading result at 8K+ GPU scale",true,"llama31_405b entries in the v6.0 training times, shortest official-style time",leading result at 8K+ GPU scale,"order: Azure 8192 GPUs 7.07 min, CoreWeave 4096 GPUs 9.77 min, NVIDIA 5120 GPUs 9.99 min, NVIDIA 2048 GPUs 18.51 min, NVIDIA 2560 GPUs 18.79 min. Azure ranks 1 of 9",CONSISTENT,"Leading is read as the shortest time, and separately as the largest GPU count.",false,"Azure: Inside Llama 3.1 405B MLPerf Training on Azure: System-Level Insights at 8K+ GPU Scale (post dated 24 June 2026, read from a republished copy on the Azure Feeds listing)",,2026-10-28,MLPerf Training v6.0
+AT3,Azure,Azure Training v6.0,"Training time drops from 18.79 minutes to 7.07 minutes, representing 83% scaling efficiency relative to the 40-rack baseline as shown in the table below.",true,arithmetic on the post's table: 18.79 over 7.07 divided by 128 over 40,83% scaling efficiency,"2.658x faster on 3.2x the racks, efficiency 83.1%. GPUs = racks x 64 and GBS = MiniBS x racks hold in all four rows: yes",CONSISTENT,Checked from the post's own numbers. The file holds only the 128-rack run.,false,"Azure: Inside Llama 3.1 405B MLPerf Training on Azure: System-Level Insights at 8K+ GPU Scale (post dated 24 June 2026, read from a republished copy on the Azure Feeds listing)",,2026-10-28,MLPerf Training v6.0
+AT5,Azure,Azure Training v6.0,"Roughly 57% of the training iteration wall time (~733 ms) is spent in compute. GEMM kernels dominate, accounting for 74% of that compute time. The FFN up-projection and down-projection GEMMs alone consume ~54% of total compute (~397 ms).",true,arithmetic on the post's step time of 1.27 s and the times it gives,"57% compute, 54% FFN of compute","733 ms over 1,270 ms = 57.7%. 397 over 733 = 54.2%",CONSISTENT,Checked from the post's own numbers. The 57% is 57.7% by the post's times.,false,"Azure: Inside Llama 3.1 405B MLPerf Training on Azure: System-Level Insights at 8K+ GPU Scale (post dated 24 June 2026, read from a republished copy on the Azure Feeds listing)",,2026-10-28,MLPerf Training v6.0
+AT4,Azure,Azure Training v6.0,"The scaling table in the post lists four scale points: 40 racks with 2,560 GPUs in 18.79 minutes, 64 racks with 4,096 GPUs in 12.86 minutes, 112 racks with 7,168 GPUs in 7.75 minutes, and 128 racks with 8,192 GPUs in 7.07 minutes.",false,Azure llama31_405b entries in the v6.0 training times,"four runs with times 18.79, 12.86, 7.75 and 7.07 minutes","1 Azure entry in the file, at 8192 GPUs",CANNOT_TELL,"The 40, 64 and 112 rack points are not Azure entries in the training result files. Only the 8,192 GPU run is. The file has an NVIDIA entry at 2,560 GPUs with 18.79 minutes (system hsg_ngpu2560), the same time the post gives for 2,560 Azure GPUs.",false,"Azure: Inside Llama 3.1 405B MLPerf Training on Azure: System-Level Insights at 8K+ GPU Scale (post dated 24 June 2026, read from a republished copy on the Azure Feeds listing)",,2026-10-28,MLPerf Training v6.0
+AT6,Azure,Azure Training v6.0,"Llama 3.1, with 405 billion active parameters, has the highest active parameter count among all models in the MLPerf suite.",true,model parameter counts are not fields in the training result files,highest active parameter count in the suite,no parameter field in the files,CANNOT_TELL,,false,"Azure: Inside Llama 3.1 405B MLPerf Training on Azure: System-Level Insights at 8K+ GPU Scale (post dated 24 June 2026, read from a republished copy on the Azure Feeds listing)",,2026-10-28,MLPerf Training v6.0
+AR2,AMD,AMD Repro v6.0,Llama 2 70B WMXFP4 MI355X 87 ... gpt-oss-120b WMXFP4 MI355X 94,true,"AMD multi-node entries 6.0-0001 and 6.0-0003 in the v6.0 file, Total Accelerators field",87 GPUs for Llama 2 70B and 94 GPUs for gpt-oss-120b,file lists 88 accelerators for Llama 2 70B and 96 for gpt-oss-120b,ABOVE,"The file counts whole platforms, 11 x 8 and 12 x 8. The post says some GPUs were unhealthy and not used, so it gives 87 and 94.",false,AMD ROCm blog: Reproducing the AMD MLPerf Inference v6.0 Submission Result (post dated 1 April 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.0-repro/README.html,2026-10-28,MLPerf Inference v6.0
+AR1,AMD,AMD Repro v6.0,Nine AMD partners submitted results on Instinct platforms in the “Available” category.,true,"v6.0 entries on AMD Instinct from submitters other than AMD, available category, combined submitter names split into organizations",nine partners,"10 submitter names (Cisco, Dell, Dell_MangoBoost, GigaComputing, HPE, MangoBoost, MiTAC, ORACLE, Supermicro and Supermicro_RedHat) which split into 9 organizations: Cisco, Dell, GigaComputing, HPE, MangoBoost, MiTAC, Oracle, RedHat and Supermicro",CONSISTENT,Our rule: a combined name such as Dell_MangoBoost or Supermicro_RedHat counts as two organizations. The file has 10 submitter names.,false,AMD ROCm blog: Reproducing the AMD MLPerf Inference v6.0 Submission Result (post dated 1 April 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.0-repro/README.html,2026-10-28,MLPerf Inference v6.0
+AR4,AMD,AMD Repro v6.0,"For multi-node submissions, 11 or 12 systems are required.",true,"AMD multi-node entries in the v6.0 file, accelerators divided by 8",11 or 12 systems,11 and 12 platforms,CONSISTENT,,false,AMD ROCm blog: Reproducing the AMD MLPerf Inference v6.0 Submission Result (post dated 1 April 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.0-repro/README.html,2026-10-28,MLPerf Inference v6.0
+AR5,AMD,AMD Repro v6.0,MLPerf Inference v6.0 marked AMD’s fourth round of submissions to MLPerf Inference.,true,"rounds with an AMD submission: the v5.0, v5.1 and v6.0 result files and the v3.1, v4.0 and v4.1 repositories",fourth round,"AMD appears in 4.1, 5.0, 5.1 and 6.0",CONSISTENT,"Earlier rounds have no summary file, so the closed division system description files in each repository were read. AMD sorts first in the file list, so the cut-short lists for v3.1 and v4.0 would show it. The open division of those two rounds is not covered.",false,AMD ROCm blog: Reproducing the AMD MLPerf Inference v6.0 Submission Result (post dated 1 April 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.0-repro/README.html,2026-10-28,MLPerf Inference v6.0
+AR6,AMD,AMD Repro v6.0,Samples per second: 365.738 Tokens per second: 103480,true,"AMD entry 6.0-0002, Llama 2 70B 99 Offline",103480 tokens per second in the sample output,"103,480 tokens/s, Tokens/s",CONSISTENT,The sample output is from a run on the reader's system and the post says results depend on the system. It equals the AMD entry. The samples per second figure is not a file field.,false,AMD ROCm blog: Reproducing the AMD MLPerf Inference v6.0 Submission Result (post dated 1 April 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.0-repro/README.html,2026-10-28,MLPerf Inference v6.0
+AR3,AMD,AMD Repro v6.0,"Each MI355X platform comes with 8 MI355X GPUs. In the clusters where the multi-node benchmarks were run, some GPUs were not in a healthy state. Therefore, the submissions did not use all GPUs in the cluster.",true,"AMD multi-node entries in the v6.0 file, accelerators divided by 8","whole platforms of 8, fewer GPUs used",88 = 11 x 8 and 96 = 12 x 8 in the file,CANNOT_TELL,"The file records the platform size, not how many GPUs ran. Whether fewer GPUs were used is not a file field.",false,AMD ROCm blog: Reproducing the AMD MLPerf Inference v6.0 Submission Result (post dated 1 April 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.0-repro/README.html,2026-10-28,MLPerf Inference v6.0
+CM1,Cisco,Cisco Blog v6.1,We are proud that it is the industry’s first multi-vendor accelerator submission,true,"entries whose accelerator field names more than one accelerator in the v5.0, v5.1, v6.0 and v6.1 files",first multi-vendor accelerator submission,v6.1 entries with mixed accelerators: Cisco Cisco UCS H200 x8 + MI350X x8 G200 2x2 TOR. Earlier rounds in the files: 0,CONSISTENT,Rounds before v5.0 are not in the files.,false,Cisco Blogs: Industry's first MLPerf Inference benchmarking across multi-vendor accelerators (post dated 21 September 2026),https://blogs.cisco.com/datacenter/industrys-first-mlperf-inference-benchmarking-across-multi-vendor-accelerators/,2026-10-28,MLPerf Inference v6.1
+CM2,Cisco,Cisco Blog v6.1,Cisco connected eight NVIDIA H200 and eight AMD Instinct,true,"Cisco entry in the v6.1 file, accelerator fields",eight H200 and eight MI350X,16 accelerators: NVIDIA H200-SXM-141GB + AMD Instinct MI350X 288GB HBM3e. Framework vLLM,CONSISTENT,The file does not give the split between the two kinds. 16 is the sum.,false,Cisco Blogs: Industry's first MLPerf Inference benchmarking across multi-vendor accelerators (post dated 21 September 2026),https://blogs.cisco.com/datacenter/industrys-first-mlperf-inference-benchmarking-across-multi-vendor-accelerators/,2026-10-28,MLPerf Inference v6.1
+CM3,Cisco,Cisco Blog v6.1,"Running the systems together produced 73,357.62 tokens per second.",true,"Cisco heterogeneous entry, GPT-OSS-120B","73,357.62 tokens/s","Server 73,357.60, Offline 94,249.30",CONSISTENT,The figure is the Server result. The file prints 73357.6.,false,Cisco Blogs: Industry's first MLPerf Inference benchmarking across multi-vendor accelerators (post dated 21 September 2026),https://blogs.cisco.com/datacenter/industrys-first-mlperf-inference-benchmarking-across-multi-vendor-accelerators/,2026-10-28,MLPerf Inference v6.1
+CM4,Cisco,Cisco Blog v6.1,"That was 104.7% of the baseline, an increase of 3,272.14 tokens per second.",true,"arithmetic on the post's 73,357.62 and 70,085.48","104.7% and +3,272.14","73,357.62 over 70,085.48 = 104.67%, difference 3,272.14",CONSISTENT,Checked from the post's own numbers.,false,Cisco Blogs: Industry's first MLPerf Inference benchmarking across multi-vendor accelerators (post dated 21 September 2026),https://blogs.cisco.com/datacenter/industrys-first-mlperf-inference-benchmarking-across-multi-vendor-accelerators/,2026-10-28,MLPerf Inference v6.1
+CM5,Cisco,Cisco Blog v6.1,"we measured each system separately and added their throughput rates to establish a 70,085.48-token-per-second baseline",true,Cisco GPT-OSS-120B Server entries in the v6.1 file,"a 70,085.48 tokens/s baseline from two standalone measurements","Cisco GPT-OSS-120B Server entries: Cisco UCS C845A M8 (8x RTXPro6000-PCIe-96GB, TensorRT) 14,817.4, Cisco_UCS_B300x16_G200_4x2_RO 214,774.0, Cisco UCS H200 x8 + MI350X x8 G200 2x2 TOR 73,357.6, Cisco UCS X210c M8 - X580P PCIe Node (4x RTXPro6000-PCIe-96GB, TensorRT) 6,950.8",CANNOT_TELL,The two standalone measurements are not entries in the file. Only the combined run is.,false,Cisco Blogs: Industry's first MLPerf Inference benchmarking across multi-vendor accelerators (post dated 21 September 2026),https://blogs.cisco.com/datacenter/industrys-first-mlperf-inference-benchmarking-across-multi-vendor-accelerators/,2026-10-28,MLPerf Inference v6.1
+CM6,Cisco,Cisco Blog v6.1,"In the GPT-OSS 120B Server run, the NVIDIA H200 group handled prefill while the AMD Instinct",true,Cisco heterogeneous entry in the v6.1 file,"H200 handled prefill, MI350X handled decode",notes field: empty. Framework: vLLM,CANNOT_TELL,Which group ran which phase is not a file field.,false,Cisco Blogs: Industry's first MLPerf Inference benchmarking across multi-vendor accelerators (post dated 21 September 2026),https://blogs.cisco.com/datacenter/industrys-first-mlperf-inference-benchmarking-across-multi-vendor-accelerators/,2026-10-28,MLPerf Inference v6.1
+AX12,AMD,AMD Tech v6.1,Here AMD Instinct MI355X GPU achieves leadership in Offline and Interactive categories against B200 and in Interactive vs B300. All other performance tests are tied (we define a tie as being within 3%).,true,"AMD eight-GPU Llama 2 70B 99.9 entry against the NVIDIA eight-GPU B200 and B300 entries, with a 3% tie rule","leadership in Offline and Interactive vs B200, Interactive vs B300, all else tied","Offline: AMD 106,517, B200 102,703 (ahead), B300 109,277 (tie). Server: AMD 103,275, B200 102,398 (tie), B300 106,949 (behind). Interactive: AMD 73,192, B200 0 (no entry), B300 0 (no entry)",BELOW,"Our reading: ahead means more than 3% higher, tie means within 3% either way. The Offline and Server comparisons do not all agree with the post.",false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX19,AMD,AMD Tech v6.1,"The Performance Results table lists entry 6.1-0109 with dlrm-v3 on 8 MI355X at 16,194.40 queries/s Server and 16,363.00 samples/s Offline.",false,"AMD DLRM-v3 entries in the v6.1 file, division field, and the closed datacenter tab of the summary spreadsheet","two closed division results of 16,194.40 and 16,363.00","file: AMD eight-GPU MI355X DLRM-v3 Server 16,194.40 and Offline 16,363.00 are in the open division. The closed division entry for the same hardware is Server 12,198.80 and Offline 13,027.70 (6.1-0003). The ID 6.1-0109 is not in the closed tab",BELOW,The post says AMD participated in the closed division. The values match the open division entry.,false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX5,AMD,AMD Tech v6.1,"The Generational Improvement table cites entry 6.0-0003 for the v6.0 scores, 6.1-0004 for the v6.1 Llama 2 70B and GPT-OSS-120B scores, and 6.0-0102 for the v6.0 Wan score.",false,entry IDs of the eight-GPU MI355X entries in the v6.0 file and the v6.1 spreadsheet,"6.0-0003 and 6.1-0004 for the eight-GPU rows, 6.0-0102 for Wan",eight-GPU entry in v6.0 is 6.0-0002 and 6.0-0003 is the 96-GPU entry. eight-GPU entry in v6.1 is 6.1-0003 and 6.1-0004 is the 72-GPU GPT-OSS-120B entry with no Llama 2 70B row. The Wan Single Stream v6.0 entry is 6.0-0103,BELOW,The values in the table match the eight-GPU entries. The IDs point to the multi-GPU entries.,false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX8,AMD,AMD Tech v6.1,"The partner submissions table in the post lists fourteen entries by ID, submitter, GPU, GPU count and benchmarks, from 6.1-0037 to 6.1-0099.",false,"v6.1 summary spreadsheet, submitter, accelerator, accelerator count and the models with results for each entry ID",14 entries with the listed benchmarks,"submitter, accelerator and count match for 14 of 14 entries. Benchmarks listed in the post with no result in the spreadsheet: 6.1-0037 (deepseek-r1), 6.1-0040 (deepseek-r1), 6.1-0050 (deepseek-r1), 6.1-0084 (deepseek-r1) and 6.1-0099 (deepseek-r1)",BELOW,"DeepSeek-R1 is listed for 6.1-0037, 6.1-0040, 6.1-0050, 6.1-0084 and 6.1-0099. In the v6.1 file the only AMD Instinct DeepSeek-R1 entry is the Crusoe 512-GPU system.",false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX1,AMD,AMD Tech v6.1,Seven partners submitted results across multiple AMD Instinct GPUs that closely tracked results by AMD,true,"v6.1 file entries on AMD Instinct accelerators from submitters other than AMD, combined names split","seven partners: Crusoe, Dell, HPE, MangoBoost, MiTAC, Oracle and Supermicro","organizations with AMD Instinct entries: Cisco, Crusoe, Dell, HPE, MangoBoost, MiTAC, Oracle and Supermicro",CONSISTENT,"All seven named partners have AMD Instinct entries. Cisco also has one, a mixed H200 and MI350X system, and is not named in the post.",false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX10,AMD,AMD Tech v6.1,"In the GPT-OSS-120B test, AMD Instinct MI355X GPU is leading both Nvidia B200 and B300 across Offline and Server scenarios",true,AMD eight-GPU GPT-OSS-120B entry against the NVIDIA eight-GPU B200 and B300 entries,ahead of B200 and B300 in Offline and Server,"Offline: AMD 121,818, NVIDIA B200 91,487, B300 106,637. Server: AMD 113,241, NVIDIA B200 90,205, B300 100,630",CONSISTENT,"NVIDIA submissions are used where they exist, the best partner entry otherwise, as the post states.",false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX11,AMD,AMD Tech v6.1,AMD maintains its lead and is ahead by 18% in Offline and 7% in Server scenarios,true,AMD 72-GPU GPT-OSS-120B entry against the NVIDIA 72-GPU GB200 entry,"+18% Offline, +7% Server","Offline 1,039,900 against 883,761 (+17.7%). Server 964,468 against 899,005 (+7.3%)",CONSISTENT,The NVIDIA GB300 72-GPU entry is higher than the AMD value.,false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX13,AMD,AMD Tech v6.1,"In this benchmark, AMD is the only submitter. To facilitate a comparison, we use the previous round, 6.0, in which Nvidia submitted DLRM v3 scores on 8xB200 GPUs. The comparison shows AMD Instinct MI355X GPU leadership by over 20%.",true,v6.1 DLRM-v3 entries by organization. AMD eight-GPU MI355X entry against NVIDIA DLRM-v3 entries in v6.0,"AMD the only submitter, over 20% ahead of NVIDIA v6.0","v6.1 organizations with DLRM-v3: AMD and Dell_AMD. AMD Server 12,199, Offline 13,028. NVIDIA v6.0 eight-B200 Server 10,007, Offline 10,547. +21.9% and +23.5%",CONSISTENT,The Dell and AMD joint MI350P entry is also in the v6.1 DLRM-v3 results. AMD is the only submitter in the sense of the post if that joint entry is counted as AMD.,false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX14,AMD,AMD Tech v6.1,AMD Instinct MI350P GPU leads the competition by a remarkable 211% and 176%,true,Dell and AMD MI350P eight-GPU GPT-OSS-120B entry against the best eight-GPU RTX PRO 6000 entry,+211% and +176%,"Offline 48,971 against 15,738 (+211.2%). Server 40,868 against 14,817 (+175.8%)",CONSISTENT,The post does not say which percentage is Offline. The two values match in the order 211% Offline and 176% Server.,false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX15,AMD,AMD Tech v6.1,"With results on llama2-70b, gpt-oss-120b, WAN text-to-video, and DLRM benchmarks (submissions by AMD and MiTAC), it delivers about 80% of the performance of AMD Instinct MI355X GPU",true,"MI350X entries (AMD 6.1-0001 and 6.1-0002, MiTAC G8825Z5) against MI355X entries (AMD, MiTAC G4826), same model and scenario",about 80% of MI355X,"82.2%. 80.4%. 80.1%. 81.3%. 76.2%. 83.6%. 78.3% (MiTAC GPT-OSS Offline and Server, Llama 2 70B Offline and Server, AMD DLRM-v3 Server and Offline, Wan SingleStream)",CONSISTENT,Our rule: about 80% is read as 70 to 90%. The DLRM-v3 ratios and the Wan one are listed. The MiTAC pair is two different vendors systems with the same model.,false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX16,AMD,AMD Tech v6.1,"GPT-OSS-120b Offline score of 5.75M/s tokens is the absolute highest token throughput ever achieved in an MLPerf submission. ... throughput of 2.90M and 2.41M for Offline and Server scenarios, respectively.",true,Crusoe 512-GPU entries against the largest tokens-per-second values in the v5.0 to v6.1 files,"5.75M tokens/s the highest ever, DeepSeek-R1 2.90M and 2.41M the highest","Crusoe GPT-OSS-120b Offline 5,749,440. Largest in v6.1 5,749,440 (Crusoe). Largest in v5.0 to v6.0 2,494,310. DeepSeek-R1 Crusoe 2,901,950 and 2,405,310, largest other DeepSeek-R1 2,705,130",CONSISTENT,Rounds before v5.0 are not in the summary files.,false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX2,AMD,AMD Tech v6.1,The results are within a few percent of AMD own scores.,true,"eight-GPU MI355X partner entries (Dell, HPE, MiTAC, Oracle, Supermicro) against the AMD eight-GPU MI355X entry, GPT-OSS-120B and Llama 2 70B 99.9, Offline and Server",within a few percent of AMD,"20 comparisons, from 94.9% to 101.2% of the AMD value. Mean 97.2%",CONSISTENT,"Our rule: a few percent is read as up to 5 whole percent, with each difference rounded to a whole percent. The lowest comparison is 5.1% below.",false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX3,AMD,AMD Tech v6.1,"The Performance Results table in the post lists 31 values by entry ID, benchmark, scenario, GPU and GPU count for AMD, Dell and AMD, and Crusoe entries.",false,"v6.1 summary spreadsheet, entry ID and result columns",31 values with entry IDs,31 of 31 match the spreadsheet to the digits printed. Mismatches none. Not found none,CONSISTENT,"The post also lists 6.1-0109 for an eight-GPU MI355X DLRM-v3 entry with 16,194.40 Server and 16,363.00 Offline. That ID is not in the spreadsheet. See the next row.",false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX4,AMD,AMD Tech v6.1,"The Generational Improvement table in the post lists v6.0 and v6.1 scores on the MI355X for Llama 2 70B Offline 103,480 to 106,517 (2.9%) and Server 100,282.36 to 103,275 (3%), GPT-OSS-120B Offline 95,004 to 121,818 (28%) and Server 82,136.10 to 113,241 (38%), and Wan 2.2 Single Stream 27.38 s to 16.124 s (70%).",false,"AMD eight-GPU MI355X entries in v6.0 (6.0-0002) and v6.1 (6.1-0003), same model and scenario","2.9%, 3%, 28%, 38%, 70%","+2.9%, +3.0%, +28.2%, +37.9%, +69.8%. v6.0 scores 103,480, 100,282.00, 95,004, 82,136.1, Wan 27.384519297 s",CONSISTENT,The Wan figure is the speed gain (27.38 s over 16.124 s). The time falls 41%.,false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX6,AMD,AMD Tech v6.1,"The scaling table in the post lists GPT-OSS 120B on 8 GPUs (6.1-0003) at 121,818 Offline and 113,241 Server, on 72 GPUs (6.1-0004) at 1,039,900 and 964,468, an ideal 72-GPU value of 1,096,362 and 1,019,169, and a scaling efficiency of 94.9% and 94.6%.",false,"AMD eight-GPU and 72-GPU GPT-OSS-120B entries, 9 times the eight-GPU value",94.9% Offline and 94.6% Server,"1,039,900 over 1,096,362 = 94.9%, 964,468 over 1,019,169 = 94.6%",CONSISTENT,,false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX7,AMD,AMD Tech v6.1,"Fifth MLPerf appearance: Fifth consecutive MLPerf Inference for AMD round reflects our sustained investment in transparent, standards-based benchmarking.",true,rounds with an AMD closed division submission in the v3.1 to v6.1 sources,fifth consecutive round,"AMD appears in 4.1, 5.0, 5.1, 6.0 and 6.1",CONSISTENT,"Rounds v3.1 and v4.0 were read from system description files with a cut-short list, where AMD would sort first.",false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX9,AMD,AMD Tech v6.1,"AMD Instinct MI355X GPU leading Nvidia B300 by 18% and 11%, respectively, for Offline and SingleStream scenarios",true,AMD eight-GPU Wan 2.2 entry against the NVIDIA eight-GPU B300 entry,"+18% Offline, +11% SingleStream",Offline 0.0731 against 0.0622 (+17.6%). SingleStream 16.124 s against 17.977 s (+11.5% faster),CONSISTENT,Another B300 entry from a partner is above the AMD Offline value.,false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX17,AMD,AMD Tech v6.1,a joint MangoBoost/Dell submission (6.1-0041) employs two different types of AMD Instinct GPUs ... they achieve a remarkably high scaling efficiency of 97% for Offline and 94% for Server in GPT-OSS-120B,true,"entry 6.1-0041 in the v6.1 summary spreadsheet, system name and accelerators",scaling efficiency 97% Offline and 94% Server,"Heterogeneous Four-Cluster (2x8x MI300X + 2x8x MI355X, LLMBoost). 32 accelerators. Offline 285,454, Server 253,501. The standalone entries it is scaled against are not named",CANNOT_TELL,The efficiency needs the standalone measurements of each cluster. They are not entries in the file. The locations are not file fields.,false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+AX18,AMD,AMD Tech v6.1,yielding an 8% end-to-end improvement in Offline and a 5% improvement in the Server scenario,true,optimization increments are not result file fields,8% Offline and 5% Server from MoE kernel tuning,no entry isolates a single optimization,CANNOT_TELL,"The same applies to the other optimization gains in the post: attention 25 to 28%, scheduler 8%, DeepSeek-R1 steps of 4 to 15%.",false,ROCm Blogs: Technical Dive into AMD MLPerf Inference v6.1 Submission (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf-v6.1/README.html,2026-10-28,MLPerf Inference v6.1
+RP1,AMD,AMD Repro v6.1,"The Expected Results table in the post lists approximate values for dlrm-v3, gpt-oss-120b, llama2-70b and wan2.2-t2v on the MI355X, MI350X and MI350P, and says runs usually vary within 3%.",false,"v6.1 summary spreadsheet, AMD and Dell and AMD entries (6.1-0001 to 6.1-0004 and 6.1-0040), same model, platform and scenario","expected values within 3% of a run, which the post gives as the usual variation","14 of 24 file values are within 3% of the expected value. Outside 3%: dlrm-v3 MI350X Offline expected 10,500, file 10,891.1 (+3.7%). dlrm-v3 MI350P Offline expected 4,500, file 4,837.63 (+7.5%). dlrm-v3 MI350P Server expected 4,200, file 4,503.53 (+7.2%). gpt-oss-120b 72 GPUs MI355X Offline expected 1,000,900, file 1,039,900 (+3.9%). llama2-70b MI350P Interactive expected 21,000, file 22,860.5 (+8.9%). llama2-70b MI350P Offline expected 41,000, file 42,286.2 (+3.1%). llama2-70b MI350P Server expected 40,000, file 41,830 (+4.6%). wan2.2-t2v MI355X Offline expected 0.07, file 0.0731282 (+4.5%). wan2.2-t2v MI350X Offline expected 0.05, file 0.0565013 (+13.0%). wan2.2-t2v MI350X Single Stream expected 20, file 20.60302473 (+3.0%)",BELOW,The 72-GPU rows are labelled as such in the post. The post does not say the expected values are targets for the submitted runs. Our rule: the 3% in the post is the tolerance.,false,ROCm Blogs: Reproducing AMD MLPerf Inference v6.1 Submission Results (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.1-repro/README.html,2026-10-28,MLPerf Inference v6.1
+RP2,AMD,AMD Repro v6.1,"The Expected Results table lists llama3.1-8b on the MI355X at about 165,000 tokens/s Offline, 155,000 Server and 140,000 Interactive.",false,eight-GPU MI355X Llama 3.1 8B entries from partners in the v6.1 file. AMD has no Llama 3.1 8B MI355X entry,"~165,000, ~155,000 and ~140,000 tokens/s","Offline expected 165,000: partner range 144,321 to 159,278. Server expected 155,000: partner range 148,118 to 151,116. Interactive expected 140,000: partner range 130,540 to 136,999",BELOW,AMD itself has no Llama 3.1 8B entry on the MI355X. The expected values are compared with the five partner entries.,false,ROCm Blogs: Reproducing AMD MLPerf Inference v6.1 Submission Results (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.1-repro/README.html,2026-10-28,MLPerf Inference v6.1
+RP4,AMD,AMD Repro v6.1,"The Workloads Summary table lists llama2-70b on MI355X and MI350P, deepseek-r1 on MI355X, gpt-oss-120b on MI355X and MI350P, dlrm-v3 on MI355X, MI350X and MI350P, wan2.2-t2v on MI355X, MI350X and MI350P, and llama3.1-8b on MI355X.",false,models and platforms of AMD and Dell and AMD entries in the v6.1 file,six models with the platforms listed,"dlrm-v3: MI350P, MI350X and MI355X, gpt-oss-120b: MI350P and MI355X, llama2-70b-99.9: MI350P and MI355X, llama3.1-8b: MI350P, wan-2.2-t2v-a14b: MI350P, MI350X and MI355X",BELOW,DeepSeek-R1 has no AMD-labelled entry. Llama 3.1 8B appears for the MI350P and not for the MI355X. The post lists the opposite for both.,false,ROCm Blogs: Reproducing AMD MLPerf Inference v6.1 Submission Results (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.1-repro/README.html,2026-10-28,MLPerf Inference v6.1
+RP5,AMD,AMD Repro v6.1,Completed samples per second : 12198.83,true,"AMD eight-GPU MI355X DLRM-v3 Server, closed division",12198.83 completed samples per second,"12,198.80 queries/s",CONSISTENT,The post prints the sample run output. The 99th percentile latency of 60.43 ms is not a file field.,false,ROCm Blogs: Reproducing AMD MLPerf Inference v6.1 Submission Results (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.1-repro/README.html,2026-10-28,MLPerf Inference v6.1
+RP6,AMD,AMD Repro v6.1,AMD MLPerf Inference v6.1 submission is our fifth consecutive round of MLPerf Inference participation.,true,rounds with an AMD submission in the v3.1 to v6.1 sources,fifth consecutive round,"AMD appears in 4.1, 5.0, 5.1, 6.0 and 6.1",CONSISTENT,"Rounds v3.1 and v4.0 were read from system description files with a cut-short list, where AMD would sort first.",false,ROCm Blogs: Reproducing AMD MLPerf Inference v6.1 Submission Results (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.1-repro/README.html,2026-10-28,MLPerf Inference v6.1
+RP3,AMD,AMD Repro v6.1,"The Expected Results table lists deepseek-r1 on the MI355X at about 51,000 tokens/s Offline and 41,000 Server.",false,AMD and Dell and AMD DeepSeek-R1 entries in the v6.1 file,"~51,000 and ~41,000 tokens/s on MI355X","AMD-labelled DeepSeek-R1 entries: 0. The only AMD Instinct DeepSeek-R1 entry is Crusoe on 512 MI355X, Offline 2,901,950",CANNOT_TELL,There is no eight-GPU MI355X DeepSeek-R1 entry in the file to compare with.,false,ROCm Blogs: Reproducing AMD MLPerf Inference v6.1 Submission Results (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.1-repro/README.html,2026-10-28,MLPerf Inference v6.1
+RP7,AMD,AMD Repro v6.1,"ROCm: [PLACEHOLDER: ROCm version, e.g., ROCm 6.4]",true,software fields of the AMD entries in the v6.1 file,a ROCm version for the prerequisites,"framework and notes fields of AMD-labelled entries name: ROCm 7.13.0, ROCm 7.2.2, ROCm 7.2.3 and rocm7.2.3",CANNOT_TELL,The published page carries the placeholder text in place of a version. The file names the versions used in the submitted runs.,false,ROCm Blogs: Reproducing AMD MLPerf Inference v6.1 Submission Results (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.1-repro/README.html,2026-10-28,MLPerf Inference v6.1
+RP8,AMD,AMD Repro v6.1,"Total disk space required across all workloads is approximately 5GB. ... Note that dlrm-v3 dominates the total at roughly 1.2 TB on its own, and also requires about 3 TB of host system memory.",true,arithmetic on the two figures in the post,5GB in total and 1.2 TB for one workload,1.2 TB is about 240 times 5GB,CANNOT_TELL,The two figures in the post cannot both be the total. Neither is a result file field.,false,ROCm Blogs: Reproducing AMD MLPerf Inference v6.1 Submission Results (post dated 17 September 2026),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inf_v6.1-repro/README.html,2026-10-28,MLPerf Inference v6.1
+RF2,AMD,AMD Repro v5.1,"Momentum across the ecosystem remains strong, with eight AMD partners submitting results on Instinct platforms. Every submission falls into the “available” category",true,"v5.1 entries on AMD Instinct accelerators from submitters other than AMD, combined names split. Availability field","eight partners, all available","9 organizations: ASUSTeK, Dell, GigaComputing, Krai, MangoBoost, MiTAC, QCT, Supermicro and Vultr. Availability values available",BELOW,Our rule: a combined name such as Dell_MangoBoost counts as two organizations. The count is nine with Krai and eight without it. Availability is available for every entry.,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.1 Submission (post dated 9 September 2025),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference5.1-repro/README.html,2026-10-28,MLPerf Inference v5.1
+RF5,AMD,AMD Repro v5.1,Mixtral 8x7B is a mixture-of-experts model making its MLPerf debut.,true,Mixtral 8x7B entries in the v5.0 file. Mixtral files in the v4.1 repository,Mixtral 8x7B debuts in v5.1,v5.0 has 13 submitters with Mixtral 8x7B entries. The v4.1 repository holds Mixtral files. No AMD Mixtral entry before v5.1,BELOW,The benchmark is older than this round. It is the first AMD Mixtral submission.,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.1 Submission (post dated 9 September 2025),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference5.1-repro/README.html,2026-10-28,MLPerf Inference v5.1
+RF1,AMD,AMD Repro v5.1,MLPerf Inference v5.1 marks AMD’s third round of submissions,true,rounds with an AMD submission in the v3.1 to v5.1 sources,third round,"AMD appears in 4.1, 5.0 and 5.1",CONSISTENT,"AMD appears first in v4.1 in the repository system files. Rounds v3.1 and v4.0 were read from lists cut short at the end, where AMD would sort first.",false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.1 Submission (post dated 9 September 2025),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference5.1-repro/README.html,2026-10-28,MLPerf Inference v5.1
+RF3,AMD,AMD Repro v5.1,"For the first time, AMD is competing in the Open division",true,"AMD open division entries in the v5.0 and v5.1 files. AMD in the v3.1, v4.0 and v4.1 repositories",first open division entries,v5.1 AMD open entries: 4. v5.0: 0. v4.1 AMD open files: 0. AMD in v3.1 and v4.0: absent,CONSISTENT,The v4.1 repository has no open division files for AMD.,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.1 Submission (post dated 9 September 2025),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference5.1-repro/README.html,2026-10-28,MLPerf Inference v5.1
+RF4,AMD,AMD Repro v5.1,"The MLPerf Inference v5.1 Submission Overview table in the post lists nine rows by model, datatype, Instinct platform and division: Llama 2 70B, Llama 2 70B interactive, Mixtral 8x7B and SD-XL on the MI325X in the Closed division, and Llama 3.1 405B pruned, Llama 3.1 405B pruned and fine-tuned, and Llama 2 70B on one, four and eight MI355X systems in the Open division.",false,"AMD and AMD_MangoBoost v5.1 entries: model, division, accelerator and count","nine rows: four closed on MI325X, five open on MI355X","closed MI325X: llama2-70b-99, llama2-70b-99.9, mixtral-8x7b and stable-diffusion-xl. Open MI355X: llama2-70b-99 32 GPUs (AMD_MangoBoost), llama2-70b-99 64 GPUs (AMD_MangoBoost), llama2-70b-99 8 GPUs (AMD), llama2-70b-99.9 32 GPUs (AMD_MangoBoost), llama2-70b-99.9 64 GPUs (AMD_MangoBoost), llama2-70b-99.9 8 GPUs (AMD), llama3.1-405b 8 GPUs (AMD)",CONSISTENT,The 4-node and 8-node rows are 32 and 64 GPU entries submitted as AMD_MangoBoost. The Mixtral and Llama 2 70B closed entries on the MI300X are also in the file and are not in the table.,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.1 Submission (post dated 9 September 2025),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference5.1-repro/README.html,2026-10-28,MLPerf Inference v5.1
+RF6,AMD,AMD Repro v5.1,Llama 3.1 405B - This workload was introduced in MLPerf inference v5.0,true,Llama 3.1 405B entries in the v5.0 file. The v4.1 repository,introduced in v5.0,v5.0 file: 11 submitters. v4.1 repository paths naming 405: 0,CONSISTENT,,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.1 Submission (post dated 9 September 2025),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference5.1-repro/README.html,2026-10-28,MLPerf Inference v5.1
+RF7,AMD,AMD Repro v5.1,tighter latency targets of 450 ms TTFT and 40 ms TPOT (25 tokens/s per user),true,arithmetic on the post's TPOT,25 tokens/s per user at 40 ms TPOT,1000 ms over 40 ms = 25 tokens/s,CONSISTENT,Checked from the post's own number.,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.1 Submission (post dated 9 September 2025),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference5.1-repro/README.html,2026-10-28,MLPerf Inference v5.1
+RF8,AMD,AMD Repro v5.1,Samples per second: 18 .5856 ... Completed samples per second : 16 .20 ... Tokens per second: 1942.23 ... Tokens per second: 2019.04 ... Tokens per second: 92081,true,"AMD v5.1 entries 5.1-0004, 5.1-0100, 5.1-0101 and 5.1-0098, result field",five sample outputs,"SDXL Offline: post 18.5856, file 18.5856 (+0.00%), SDXL Server completed: post 16.2, file 16.2044 (-0.03%), 405B pruned: post 1942.23, file 1942.15 (+0.00%), 405B pruned and fine-tuned: post 2019.04, file 2019.04 (+0.00%), Llama 2 70B MXFP4: post 92081, file 93045.8 (-1.04%)",CONSISTENT,Sample outputs come from a run on the reader's system and the post says results depend on the system. Our rule: within 2% of the file entry.,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.1 Submission (post dated 9 September 2025),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-inference5.1-repro/README.html,2026-10-28,MLPerf Inference v5.1
+RA1,AMD,AMD Repro v5.0,The AMD MLPerf Inference v5.0 submission includes results that were generated on a system with 8 x MI325X GPUs and 2 x AMD EPYC 9655 CPUs.,true,"AMD v5.0 entries 5.0-0001 (Llama 2 70B) and 5.0-0002 (SDXL), processor and accelerator fields",8 x MI325X and 2 x EPYC 9655,"5.0-0001 Llama 2 70B: 2 x AMD EPYC 9575F, 8 AMD Instinct MI325X 256GB HBM3E. 5.0-0002 SDXL: 2 x 2xAMD EPYC 9655, 8 AMD Instinct MI325X 256GB HBM3E",BELOW,The SDXL system has the EPYC 9655. The Llama 2 70B system in the file has the EPYC 9575F.,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.0 Submission (post dated 2 April 2025),https://rocm.blogs.amd.com/artificial-intelligence/reproducing-amd-mlperf-inference-submission/README.html,2026-10-28,MLPerf Inference v5.0
+RA5,AMD,AMD Repro v5.0,Tokens per second: 33396.9 ... Completed tokens per second: 29522.56 ... Samples per second: 17 .0963,true,"AMD v5.0 entries 5.0-0001 and 5.0-0002, result field",three sample outputs,"Llama 2 70B Offline: post 33396.9, file 33928.1 (-1.57%), Llama 2 70B Server: post 29522.6, file 30724.5 (-3.91%), SDXL Offline: post 17.0963, file 17.0963 (+0.00%)",BELOW,Sample outputs come from a run on the reader's system. Our rule: within 2% of the file entry. Llama 2 70B Server is 3.9% below.,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.0 Submission (post dated 2 April 2025),https://rocm.blogs.amd.com/artificial-intelligence/reproducing-amd-mlperf-inference-submission/README.html,2026-10-28,MLPerf Inference v5.0
+RA2,AMD,AMD Repro v5.0,"These results are submitted in the Available category, showcasing AMD hardware and software that are already available to customers.",true,availability and category fields of the AMD v5.0 entries,available category,availability available. Division closed,CONSISTENT,,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.0 Submission (post dated 2 April 2025),https://rocm.blogs.amd.com/artificial-intelligence/reproducing-amd-mlperf-inference-submission/README.html,2026-10-28,MLPerf Inference v5.0
+RA3,AMD,AMD Repro v5.0,AMD has submitted results for two popular models,true,models in the AMD v5.0 entries,two models,"models llama2-70b-99, llama2-70b-99.9 and stable-diffusion-xl",CONSISTENT,Llama 2 70B appears at both 99 and 99.9 accuracy. SDXL is the second model.,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.0 Submission (post dated 2 April 2025),https://rocm.blogs.amd.com/artificial-intelligence/reproducing-amd-mlperf-inference-submission/README.html,2026-10-28,MLPerf Inference v5.0
+RA4,AMD,AMD Repro v5.0,Building upon the success of our MLPerf Inference v4.1 submission,true,AMD in the v4.1 repository,a v4.1 submission,AMD in v4.1 repository: yes,CONSISTENT,,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.0 Submission (post dated 2 April 2025),https://rocm.blogs.amd.com/artificial-intelligence/reproducing-amd-mlperf-inference-submission/README.html,2026-10-28,MLPerf Inference v5.0
+RA6,AMD,AMD Repro v5.0,ROCm 6.3.3 or later,true,Software field of the AMD v5.0 entry 5.0-0001,ROCm 6.3.3 or later for reproduction,the submitted Llama 2 70B run lists ROCm 6.3.1. The SDXL entry lists shark-ai 3.3.0,CANNOT_TELL,The post asks for a later ROCm than the one the Llama 2 70B entry lists. The SDXL entry names no ROCm version.,false,ROCm Blogs: Reproducing the AMD Instinct GPUs MLPerf Inference v5.0 Submission (post dated 2 April 2025),https://rocm.blogs.amd.com/artificial-intelligence/reproducing-amd-mlperf-inference-submission/README.html,2026-10-28,MLPerf Inference v5.0
+PR1,AMD,AMD Pruning v5.1,pruning 26 layers (~21%) from the original 126-layer Llama 3.1 405B model ... pruning 42 layers (33.3%),true,arithmetic on the post's layer counts,~21% and 33.3% of 126 layers,"26 of 126 = 20.6%, 42 of 126 = 33.3%",CONSISTENT,Checked from the post's own numbers.,false,ROCm Blogs: Slim Down Your Llama: Pruning & Fine-Tuning for Maximum Performance (post dated 9 September 2025),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-llama-pruning/README.html,2026-10-28,MLPerf Inference v5.1
+PR2,AMD,AMD Pruning v5.1,"In its original BF16 format, the model is 756 GB. Thanks to the MI355X’s support for MXFP4 quantization, we can quantize the model to approximately 207 GB.",true,arithmetic: 405 billion parameters at 16 bits and at 4.25 bits (4 bits plus one 8-bit scale per 32 values),756 GB in BF16 and about 207 GB in MXFP4,BF16 810 GB decimal or 754 GiB. MXFP4 215 GB decimal or 200 GiB,CANNOT_TELL,Not a result file field. The BF16 figure matches the binary unit. The MXFP4 figure lies between the two computed values.,false,ROCm Blogs: Slim Down Your Llama: Pruning & Fine-Tuning for Maximum Performance (post dated 9 September 2025),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-llama-pruning/README.html,2026-10-28,MLPerf Inference v5.1
+PR3,AMD,AMD Pruning v5.1,with Exact Match accuracy above 96.7% and RougeL accuracy above 97.7% ... above 96% for Exact Match and above 99.6% for RougeL metrics,true,"accuracy fields of AMD open entries 5.1-0100 (2019.04 tokens/s) and 5.1-0101 (1942.15 tokens/s), and of the closed Llama 3.1 405B entries",accuracy percentages relative to the reference score,"5.1-0100: ROUGEL 21.170, exact match 87.22. 5.1-0101: ROUGEL 21.619, exact match 86.60. Mean of 23 closed entries: ROUGEL 21.899, exact match 90.08",CANNOT_TELL,The percentages are relative to a reference score that is not a field in the file. Against the closed-entry mean the exact match ratios are 96.8% and 96.1% and the ROUGEL ratios are 96.7% and 98.7%.,false,ROCm Blogs: Slim Down Your Llama: Pruning & Fine-Tuning for Maximum Performance (post dated 9 September 2025),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-llama-pruning/README.html,2026-10-28,MLPerf Inference v5.1
+PR4,AMD,AMD Pruning v5.1,"with generated tokens/s ranging from 2,307 to 2,921 and Exact Match accuracy from 78.35% to 88.78%",true,Table 1 experiments in the post. AMD open entries in the v5.1 file,"2,307 to 2,921 tokens/s across pruning strategies","submitted open entries: 1,942.15 and 2,019.04 tokens/s",CANNOT_TELL,Table 1 is an experiment table and not a submission. The submitted entries are lower than the experiment range.,false,ROCm Blogs: Slim Down Your Llama: Pruning & Fine-Tuning for Maximum Performance (post dated 9 September 2025),https://rocm.blogs.amd.com/artificial-intelligence/mlperf-llama-pruning/README.html,2026-10-28,MLPerf Inference v5.1
+NZ4,NVIDIA,NVIDIA v6.0,the largest number of partners submitting on any platform,true,"closed entries on NVIDIA accelerators by submitters other than NVIDIA, v5.0, v5.1 and v6.0 files","14 partners, the largest number",v6.0: 16 submitters. v5.1: 20. v5.0: 16. The 14 named partners present in v6.0: 14. Other v6.0 submitters: GATEOverflow and Krai,BELOW,"Rounds before v5.0 are not in the files. Two more submitters on NVIDIA accelerators in v6.0, GATEOverflow and Krai, are not among the 14 named. In v5.1 the combined names Broadcom_Supermicro and Dell_Broadcom split into organizations give 19, or 16 if GATEOverflow, The_Stage and University-of-Florida are left out.",false,NVIDIA Technical Blog: NVIDIA Platform Delivers Lowest Token Cost Enabled by Extreme Co-Design (post dated 1 April 2026),https://developer.nvidia.com/blog/?p=115040,2026-10-28,MLPerf Inference v6.0
+NZ1,NVIDIA,NVIDIA v6.0,"Table 1 in the post lists NVIDIA platform throughput on DeepSeek-R1 (2,494,310 Offline, 1,555,110 Server, 250,634 Interactive tokens/s), GPT-OSS-120B (1,046,150, 1,096,770 and 677,199), Qwen3-VL (79 samples/s and 68 queries/s), Wan 2.2 (0.059 samples/s and 21 s single stream) and DLRMv3 (104,637 samples/s and 99,997 queries/s).",false,"entries 6.0-0039, 6.0-0073, 6.0-0075, 6.0-0076, 6.0-0078, 6.0-0081 and 6.0-0094 in the v6.0 file, result field rounded to the digits printed",twelve values,"deepseek-r1 Offline 2,494,310. deepseek-r1 Server 1,555,110. deepseek-r1 Interactive 250,634. gpt-oss-120b Offline 1,046,150. gpt-oss-120b Server 1,096,770. gpt-oss-120b Interactive 677,199. qwen3-vl-235b-a22b Offline 79.0373. qwen3-vl-235b-a22b Server 67.8642. wan-2.2-t2v-a14b Offline 0.0585234. wan-2.2-t2v-a14b SingleStream 20.89789424. dlrm-v3 Offline 104,637. dlrm-v3 Server 99,996.8. Values not matching to the digits printed: none",CONSISTENT,"The cited entries are not all NVIDIA submissions: 6.0-0081 is Nebius, 6.0-0039 is GigaComputing and 6.0-0094 is Red Hat, all on NVIDIA accelerators. The table mixes systems of 8, 72 and 288 accelerators.",false,NVIDIA Technical Blog: NVIDIA Platform Delivers Lowest Token Cost Enabled by Extreme Co-Design (post dated 1 April 2026),https://developer.nvidia.com/blog/?p=115040,2026-10-28,MLPerf Inference v6.0
+NZ2,NVIDIA,NVIDIA v6.0,"This round, NVIDIA GB300 NVL72—launched last year—delivered up to 2.7x higher token throughput compared to its debut submissions just six months ago on the server scenario of the DeepSeek-R1 benchmark",true,"Nebius entry 6.0-0081 and NVIDIA entry 5.1-0072, DeepSeek-R1 Server, both 72 GB300","up to 2.7x, cost per token down more than 60%","575,580 over 209,328 = 2.75x. Cost per token falls 64%",CONSISTENT,The ratio is 2.75x. The cost figure is arithmetic on that ratio.,false,NVIDIA Technical Blog: NVIDIA Platform Delivers Lowest Token Cost Enabled by Extreme Co-Design (post dated 1 April 2026),https://developer.nvidia.com/blog/?p=115040,2026-10-28,MLPerf Inference v6.0
+NZ3,NVIDIA,NVIDIA v6.0,"Table 2 in the post lists GB300 NVL72 per-GPU throughput in v5.1 and v6.0: DeepSeek-R1 Server 2,907 and 8,064 tokens/s (2.77x) and Offline 5,842 and 9,821 (1.68x), and Llama 3.1 405B Server 170 and 259 (1.52x) and Offline 224 and 271 (1.21x).",false,"entries 5.1-0072, 6.0-0017, 6.0-0078 and 6.0-0082, result divided by accelerator count",eight per-GPU values and four speedups,"DeepSeek-R1 Server: 2,907 to 8,064 (2.77x, v6.0 entry on 8 accelerators). DeepSeek-R1 Offline: 5,842 to 9,821 (1.68x, v6.0 entry on 8 accelerators). Llama 3.1 405B Server: 170 to 259 (1.52x, v6.0 entry on 72 accelerators). Llama 3.1 405B Offline: 224 to 271 (1.21x, v6.0 entry on 72 accelerators)",CONSISTENT,"The values reproduce. The v5.1 entry has 72 accelerators. The v6.0 DeepSeek-R1 values come from entries with 8 accelerators (Nebius 6.0-0082 and CoreWeave 6.0-0017). At 72 accelerators the v6.0 DeepSeek-R1 Server value is 5,562 for NVIDIA 6.0-0078 and 7,994 for Nebius 6.0-0081.",false,NVIDIA Technical Blog: NVIDIA Platform Delivers Lowest Token Cost Enabled by Extreme Co-Design (post dated 1 April 2026),https://developer.nvidia.com/blog/?p=115040,2026-10-28,MLPerf Inference v6.0
+NZ5,NVIDIA,NVIDIA v6.0,With 288 Blackwell Ultra GPUs—the largest scale ever submitted to any benchmark in MLPerf Inference,true,"largest accelerator count in the v5.0, v5.1 and v6.0 files","288 GPUs, the largest scale ever",v6.0 largest 288. v5.0 and v5.1 largest 72,CONSISTENT,As of v6.0. Crusoe later submitted 512 GPUs in v6.1. Rounds before v5.0 are not in the files.,false,NVIDIA Technical Blog: NVIDIA Platform Delivers Lowest Token Cost Enabled by Extreme Co-Design (post dated 1 April 2026),https://developer.nvidia.com/blog/?p=115040,2026-10-28,MLPerf Inference v6.0
+NZ6,NVIDIA,NVIDIA v6.0,"Only the NVIDIA platform submitted results on all newly added models and scenarios this round, and delivered the highest performance across all of them.",true,"closed v6.0 entries for the five new models by accelerator vendor, and the highest result for the models where another vendor submitted","only NVIDIA on all new models, highest on all","accelerator vendors per new model: deepseek-r1 NVIDIA. qwen3-vl-235b-a22b NVIDIA. gpt-oss-120b AMD, NVIDIA and other. wan-2.2-t2v-a14b NVIDIA. dlrm-v3 NVIDIA. Highest result by vendor: gpt-oss-120b Offline NVIDIA, gpt-oss-120b Server NVIDIA, gpt-oss-120b Interactive NVIDIA, wan-2.2-t2v-a14b Offline NVIDIA",CONSISTENT,AMD submitted GPT-OSS-120B and an open Wan entry and none of the other new models. The Wan single stream open entry from AMD is not in the closed comparison.,false,NVIDIA Technical Blog: NVIDIA Platform Delivers Lowest Token Cost Enabled by Extreme Co-Design (post dated 1 April 2026),https://developer.nvidia.com/blog/?p=115040,2026-10-28,MLPerf Inference v6.0
+NZ7,NVIDIA,NVIDIA v6.0,NVIDIA was the first and only platform to submit DeepSeek-R1 results on MLPerf Inference when the benchmark debuted last year.,true,accelerator vendors of v5.1 DeepSeek-R1 entries,only NVIDIA accelerators in v5.1,vendors: NVIDIA,CONSISTENT,,false,NVIDIA Technical Blog: NVIDIA Platform Delivers Lowest Token Cost Enabled by Extreme Co-Design (post dated 1 April 2026),https://developer.nvidia.com/blog/?p=115040,2026-10-28,MLPerf Inference v6.0
+NZ8,NVIDIA,NVIDIA v6.0,was once again the only platform to submit on the newly added interactive scenario,true,accelerator vendors of v6.0 DeepSeek-R1 Interactive entries,only NVIDIA accelerators,vendors: NVIDIA,CONSISTENT,,false,NVIDIA Technical Blog: NVIDIA Platform Delivers Lowest Token Cost Enabled by Extreme Co-Design (post dated 1 April 2026),https://developer.nvidia.com/blog/?p=115040,2026-10-28,MLPerf Inference v6.0
+NZ9,NVIDIA,NVIDIA v6.0,GB300 NVL72 performance increased by 1.5x in the server scenario,true,"NVIDIA entries 5.1-0072 and 6.0-0078, Llama 3.1 405B Server per accelerator",1.5x,1.52x (170.1 to 258.7 tokens/s per GPU),CONSISTENT,,false,NVIDIA Technical Blog: NVIDIA Platform Delivers Lowest Token Cost Enabled by Extreme Co-Design (post dated 1 April 2026),https://developer.nvidia.com/blog/?p=115040,2026-10-28,MLPerf Inference v6.0
+NZ10,NVIDIA,NVIDIA v6.0,"This brings the cumulative NVIDIA MLPerf training and inference wins since 2018 to 291, which is 9x of all other submitters combined.",true,wins are not fields in the result files,"291 cumulative wins, 9x all other submitters",no win count in the files,CANNOT_TELL,A win depends on a ranking rule the post does not state.,false,NVIDIA Technical Blog: NVIDIA Platform Delivers Lowest Token Cost Enabled by Extreme Co-Design (post dated 1 April 2026),https://developer.nvidia.com/blog/?p=115040,2026-10-28,MLPerf Inference v6.0
+NO3,NVIDIA,NVIDIA v5.0,"Table 1 in the post lists, for eight B200 and eight H200 GPUs, Llama 2 70B at 98,443 Server and 98,858 Offline against 33,072 and 34,988 tokens/s (3x and 2.8x), Mixtral 8x7B at 126,845 and 128,148 against 61,802 and 62,630 (2.1x and 2.1x), and Stable Diffusion XL at 28.44 and 30.38 against 18.30 and 18.99 (1.6x and 1.6x).",false,"NVIDIA entries 5.0-0056 (8 B200) and 5.0-0060 (8 H200), Server and Offline",twelve values and six speedups,"mismatches to the digits printed: Mixtral 8x7B: B200 Server 126,845.0 and Offline 128,148.0, H200 Server 61,802.0 and Offline 62,629.6, speedups 2.05x and 2.05x. Mixtral Offline 128,148 over 62,630 = 2.046x",BELOW,"Speedups are rounded to one decimal as in the post. The Mixtral Offline speedup is 2.046x, which rounds to 2.0x. The post prints 2.1x. All other values and speedups match.",false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers Massive Performance Leaps in MLPerf Inference v5.0 (post dated 2 April 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-massive-performance-leaps-in-mlperf-inference-v5-0,2026-10-28,MLPerf Inference v5.0
+NO1,NVIDIA,NVIDIA v5.0,"On the Llama 3.1 405B benchmark, GB200 NVL72 delivered up to 3.4x higher per-GPU performance compared to the NVIDIA H200 Tensor Core eight-GPU system.",true,"NVIDIA entries 5.0-0058 (72 GB200) and 5.0-0060 (8 H200), Llama 3.1 405B, result per accelerator",3.4x Server and 2.8x Offline per GPU,"Server 3.37x, Offline 2.75x per GPU",CONSISTENT,"Per GPU is total divided by accelerator count, which the post states.",false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers Massive Performance Leaps in MLPerf Inference v5.0 (post dated 2 April 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-massive-performance-leaps-in-mlperf-inference-v5-0,2026-10-28,MLPerf Inference v5.0
+NO2,NVIDIA,NVIDIA v5.0,"At the system level, GB200 NVL72 increases performance by up to 30x, through a combination of much higher per-GPU performance as well as 9x more GPUs in the system",true,"NVIDIA entries 5.0-0058 and 5.0-0060, Llama 3.1 405B, whole system result","up to 30x, 9x more GPUs","Server 30.4x, Offline 24.8x. 72 over 8 accelerators = 9x",CONSISTENT,Offline is lower than the up to figure.,false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers Massive Performance Leaps in MLPerf Inference v5.0 (post dated 2 April 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-massive-performance-leaps-in-mlperf-inference-v5-0,2026-10-28,MLPerf Inference v5.0
+NO4,NVIDIA,NVIDIA v5.0,"On the Llama 2 70B Interactive benchmark, the eight-GPU B200 system achieved 3.1x higher throughput compared to the NVIDIA submission using eight H200 GPUs.",true,"NVIDIA entries 5.0-0056 and 5.0-0060, Llama 2 70B Interactive Server",3.1x,"59,409 over 19,424 = 3.06x",CONSISTENT,,false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers Massive Performance Leaps in MLPerf Inference v5.0 (post dated 2 April 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-massive-performance-leaps-in-mlperf-inference-v5-0,2026-10-28,MLPerf Inference v5.0
+NO5,NVIDIA,NVIDIA v5.0,The NVIDIA platform was also the only one to submit results on the Mixtral 8x7B benchmark,true,accelerator vendors of closed datacenter v5.0 Mixtral 8x7B entries,only NVIDIA accelerators,vendors: NVIDIA. Submitters: 12,CONSISTENT,"Platform is read as accelerator vendor, so partner submissions count as NVIDIA.",false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers Massive Performance Leaps in MLPerf Inference v5.0 (post dated 2 April 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-massive-performance-leaps-in-mlperf-inference-v5-0,2026-10-28,MLPerf Inference v5.0
+NO6,NVIDIA,NVIDIA v5.0,NVIDIA submitted results on every benchmark in the data center category,true,models in closed datacenter v5.0 entries and models with an entry from the NVIDIA submitter,every benchmark,models with no NVIDIA submitter entry: bert-99 and bert-99.9,CONSISTENT,"The two models without a NVIDIA entry are BERT, which the post does not list among the returning benchmarks.",false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers Massive Performance Leaps in MLPerf Inference v5.0 (post dated 2 April 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-massive-performance-leaps-in-mlperf-inference-v5-0,2026-10-28,MLPerf Inference v5.0
+NO7,NVIDIA,NVIDIA v5.0,This round adds three new benchmarks: Llama 3.1 405B ... Llama 2 70B Interactive ... Relational Graph Attention Network (R-GAT),true,"models in the v5.0 file. Paths in the v4.1 repository naming rgat, 405 or interactive",three new benchmarks in v5.0,"v5.0 models: llama2-70b-interactive-99, llama2-70b-interactive-99.9, llama3.1-405b and rgat. v4.1 repository paths naming them: 0",CONSISTENT,,false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers Massive Performance Leaps in MLPerf Inference v5.0 (post dated 2 April 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-massive-performance-leaps-in-mlperf-inference-v5-0,2026-10-28,MLPerf Inference v5.0
+NO8,NVIDIA,NVIDIA v5.0,"Additionally, NVIDIA ran the Llama 2 70B benchmark from MLPerf Inference v4.1 on GB200 NVL72, achieving an unverified result of 869,203 tokens/second.",true,a v4.1 result that is not an entry in the files,"869,203 tokens/s unverified",no v4.1 GB200 NVL72 entry in the files,CANNOT_TELL,The post marks it unverified.,false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers Massive Performance Leaps in MLPerf Inference v5.0 (post dated 2 April 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-massive-performance-leaps-in-mlperf-inference-v5-0,2026-10-28,MLPerf Inference v5.0
+NO9,NVIDIA,NVIDIA v5.0,"On the Llama 2 70B benchmark, NVIDIA H100 Tensor Core GPU throughput driven by software optimizations has increased by up to 1.5x in the last year.",true,"needs v4.1 H100 Llama 2 70B entries, which are not in the files",1.5x in a year,v4.1 H100 results are not in the files,CANNOT_TELL,,false,NVIDIA Technical Blog: NVIDIA Blackwell Delivers Massive Performance Leaps in MLPerf Inference v5.0 (post dated 2 April 2025),https://developer.nvidia.com/blog/nvidia-blackwell-delivers-massive-performance-leaps-in-mlperf-inference-v5-0,2026-10-28,MLPerf Inference v5.0
+AM1,AMD,Supp AMD v6.1,A 72-GPU GPT-OSS-120B submission exhibits excellent scalability of 95% in both Offline and Server scenarios and reaches 1 million tokens/second of performance.,true,AMD 72-GPU GPT-OSS-120B entry (9 nodes of 8 x MI355X) against 9 times the AMD single-node entry (8 x MI355X),"95% in both scenarios, 1 million tokens/s","Offline 1,039,900 against 9 x 121,818 = 94.9%. Server 964,468 against 9 x 113,241 = 94.6%",CONSISTENT,"Offline passes 1 million tokens/s. Server is 964,468.",false,"MLCommons: MLPerf Inference v6.1 Results Discussion, AMD statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+AM2,AMD,Supp AMD v6.1,expanding from three models in v6.0 to seven models across three AMD,true,closed models in v6.0 AMD entries and closed models on AMD accelerators in v6.1 from any submitter,"three models in v6.0, seven in v6.1","v6.0 AMD entries: 3 models. v6.1 entries on AMD accelerators from all submitters: 7 models (deepseek-r1, dlrm-v3, gpt-oss-120b, llama2-70b-99.9, llama3.1-8b, qwen3.6-27b, wan-2.2-t2v-a14b). AMD own v6.1 entries: 4 models",CONSISTENT,The seven count includes partner entries. AMD own entries cover four models.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, AMD statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+AM3,AMD,Supp AMD v6.1,representing the largest scale MLPerf Inference result to date,true,"largest accelerator count in rounds v5.0, v6.0 and v6.1 (the rounds we hold)","512 GPUs, the largest to date",v6.1 largest: 512. v6.0 largest: 288. v5.0 largest: 8. Older rounds not read,CONSISTENT,Rounds before v5.0 are not in our files.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, AMD statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+CR1,Crusoe,Supp Crusoe v6.1,The cluster sustained over 5.7 million tokens per second on gpt-oss-120b and over 2.9 million tokens per second on DeepSeek-R1 across Server and Offline scenarios,true,Crusoe 512-GPU entries (64 nodes of 8 x MI355X),"over 5.7 million and over 2.9 million, Server and Offline","GPT-OSS-120B Offline 5,749,440, Server 5,392,930. DeepSeek-R1 Offline 2,901,950, Server 2,405,310",BELOW,Offline passes both figures. Server is 5.39 million and 2.41 million.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Crusoe statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+CR2,Crusoe,Supp Crusoe v6.1,"The centerpiece: the first 512-GPU submissions for DeepSeek-R1 and gpt-oss-120b, run across 64 Crusoe Cloud mi355x-288gb-roce.8x nodes",true,Crusoe 512-GPU entries,"512 GPUs on 64 nodes, DeepSeek-R1 and GPT-OSS-120B","512 accelerators on 64 nodes, models deepseek-r1 and gpt-oss-120b",CONSISTENT,No earlier round we hold has a 512-GPU entry.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Crusoe statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+AS1,ASUS,Supp ASUS v6.1,"Highlights include 113,455 tokens/s in the Llama 2-70B Server scenario and 113,007",true,ASUS XA NB3I-E12 (8 x B300) Llama 2 70B 99.9 Server and Offline,"113,455 and 113,007","113,455 and 113,007",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, ASUS statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+AS2,ASUS,Supp ASUS v6.1,"more than 109,000 tokens/s on GPT-OSS-120B",true,"same entry, GPT-OSS-120B Server and Offline","more than 109,000","Offline 109,368, Server 109,949",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, ASUS statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+AS3,ASUS,Supp ASUS v6.1,"69,847 tokens/s Offline on the DeepSeek-R1 reasoning model",true,"same entry, DeepSeek-R1 Offline","69,847","69,846.9",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, ASUS statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+CI2,Cisco,Supp Cisco v6.1,achieved exact 2x linear scaling from 4 to 8 RTX Pro GPUs,true,"Cisco 4-GPU X-Series nodes (X215c Llama 3.1-8B, X210c Llama 2 70B) against the Cisco 8-GPU C845A RTX PRO 6000 entry, Offline and Server",exact 2x,"Ratios 2.007 to 2.045 (2.025, 2.007, 2.012, 2.045)",CONSISTENT,All four ratios are at or a little above 2.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Cisco statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+CI3,Cisco,Supp Cisco v6.1,"scaled DeepSeek-R1 to 137,067 tokens/s at 97.5% linear scaling efficiency over single-node performance",true,Cisco 16-GPU B300 Offline entry against two times the Cisco 8-GPU C880A B300 Offline entry,"137,067 tokens/s, 97.5%","137,067 against 2 x 70,325.4 = 97.5%",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Cisco statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+CI1,Cisco,Supp Cisco v6.1,"The deployment achieved 99.9% aggregation efficiency on Llama 2 70B (99,684 tokens/s) and 99.8% on Llama 3.1 8B (163,292 tokens/s) over standalone baselines",true,Cisco H200 x8 + MI350X x8 entry (16 GPUs) in the v6.1 file,"99.9% and 99.8%, 99,684 and 163,292 tokens/s","Entry values 99,683.8 and 163,292 tokens/s Offline. The file has no standalone H200 or MI350X entry from Cisco to divide by",CANNOT_TELL,"The two throughput values match. The efficiency needs the standalone baselines, which are not in the file.",false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Cisco statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+CW1,CoreWeave,Supp CoreWeave v6.1,"aggregate total throughput of 1.16 million tokens per second, per-GPU throughput improved 17% in Offline and 8% in Server since MLPerf v6.0",true,"CoreWeave GB300 NVL72 (72 GPUs) GPT-OSS-120B in v6.1 against entry 6.0-0016 (64 GPUs), per GPU","1.16 million tokens/s, per GPU +17% Offline and +8% Server","v6.1 Server 1,160,490, Offline 1,197,710. Per GPU 16,635 against 14,226 (+16.9%) Offline, 16,118 against 14,946 (+7.8%) Server",CONSISTENT,1.16 million is the Server value. Offline is 1.20 million.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, CoreWeave statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+FU1,Fujitsu,Supp Fujitsu v6.1,outperformed our results from the previous round (5.0-0023; NVIDIA H100-NVL-94GB x8) by approximately 50%,true,"Fujitsu v6.1 PRIMERGY CDI (8 x RTX PRO 6000) against entry 5.0-0023 (8 x H100 NVL), Llama 2 70B 99.9",approximately 50%,"Offline 26,626.5 against 18,084.9 (+47.2%). Server 27,340.9 against 16,848.6 (+62.3%)",CANNOT_TELL,"The statement names no scenario. Offline is near 50% and Server is higher, so only one of the two is approximately 50%.",false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Fujitsu statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+HP1,HPE,Supp HPE v6.1 v6.1,"126,000 tokens-per-second and a significant 7,800 tokens-per-second per GPU using sixteen GPUs on DeepSeek-R1 (open division)",true,"HPE GB200 NVL72 16-GPU DeepSeek-R1 Offline, open division","126,000 and 7,800 per GPU","126,223 and 7,889 per GPU",CONSISTENT,The statement rounds down.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, HPE statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+HP2,HPE,Supp HPE v6.1 v6.1,"101,000 tokens-per-second and 12,600 tokens-per-second per GPU using eight GPUs on GPT-OSS 20B",true,"HPE GB200 NVL72 8-GPU entry, closed division","101,000 and 12,600 per GPU, GPT-OSS 20B","Offline 101,453 (12,682 per GPU), Server 100,714 (12,589 per GPU). The model in the file is gpt-oss-120b",CONSISTENT,The statement says 20B. The file lists 120b for this entry.,true,"MLCommons: MLPerf Inference v6.1 Results Discussion, HPE statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+HP3,HPE,Supp HPE v6.1 v6.1,"achieved 136,000 tokens-per-second, and significantly, achieved an 8,500 tokens-per-second per GPU on DeepSeek-R1",true,"HPE XD690 16-GPU B300 DeepSeek-R1, closed division","136,000 and 8,500 per GPU","Offline 136,854 (8,553 per GPU). Server 80,537",CONSISTENT,The figure is the Offline result.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, HPE statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+HP4,HPE,Supp HPE v6.1 v6.1,"A single HPE Compute XD690 achieved 114,000 tokens-per-second and 14,300 tokens-per-second per GPU on GPT-OSS",true,HPE XD690 8-GPU B300 GPT-OSS-120B Offline,"114,000 and 14,300 per GPU","114,569 (14,321 per GPU)",CONSISTENT,The statement rounds down.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, HPE statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+IN2,Intel,Supp Intel v6.1,"growing from two benchmarked Xeon 6 SKUs in v6.0 to five in v6.1, and from 24 to 35 CPU inference results",true,"closed datacenter rows with no accelerator, v6.0 and v6.1","two SKUs to five, 24 to 35 results","v6.0: 24 rows, 2 processor models. v6.1: 33 rows, 5 processor models (Intel(R) Xeon(R) 6787P, Intel(R) Xeon(R) 6788P, Intel(R) Xeon(R) 6972P, Intel(R) Xeon(R) 6980P, Intel(R) Xeon(R) 6987P-C)",BELOW,"The v6.0 count and both SKU counts reproduce. The v6.1 file holds 33 such rows, not 35.",false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Intel statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+IN1,Intel,Supp Intel v6.1,"The Intel statement says that on the same silicon and socket count as v6.0, Llama 3.1-8B Server throughput on Xeon 6980P rose 2.4 times (+142%) and Offline throughput rose 56%.",false,"Intel 1-node-2S-GNR_128C, v6.0 entry 6.0-0057 against the v6.1 entry, Llama 3.1-8B","2.4x (+142%) Server, +56% Offline","Server 471.5 to 1,139.5 (2.42x, +141.7%). Offline 1,229.6 to 1,916.7 (+55.9%)",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Intel statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+IN3,Intel,Supp Intel v6.1,On this configuration gpt-oss-120B improved 36% in Server and 27% in Offline over v6.0 on the same system,true,"Intel 4 x Arc Pro B70 entry, v6.0 entry 6.0-0101 against the v6.1 entry, GPT-OSS-120B","+36% Server, +27% Offline","Server 951.7 to 1,296.9 (+36.3%). Offline 1,536.9 to 1,956.4 (+27.3%)",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Intel statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+IN4,Intel,Supp Intel v6.1,is the head node for 47 of 111 systems submitted in v6.1,true,"distinct submitter and system name pairs in the v6.1 file, and those whose host processor is a Xeon",47 of 111,47 of 116 distinct systems in the file have a Xeon host,CANNOT_TELL,"The file does not define a system. The Xeon count of 47 reproduces. The total of 111 does not. The MLCommons spreadsheet lists 120 system rows across its four tabs, 49 with a Xeon host.",false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Intel statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+MB1,MangoBoost,Supp MangoBoost v6.1,"we achieved a staggering 285,454 tokens/second in Offline mode with 97% scaling efficiency, and 253,501 tokens/second in Server mode",true,Dell MangoBoost four-cluster GPT-OSS-120B entry against 2 x the MangoBoost 8 x MI300X entry plus 2 x the Dell MangoBoost XE9785L 8 x MI355X entry (the XE9785 entry gives a second baseline),"285,454 Offline at 97% efficiency, 253,501 Server","Offline 285,454 against 294,303 = 97.0% (XE9785 baseline 98.1%). Server 253,501 against 270,032 = 93.9%",CONSISTENT,The baseline entries are our choice. The post does not name them. Server efficiency is lower.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, MangoBoost statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+NV4,NVIDIA,Supp NVIDIA v6.1,reaching a peak throughput of 0.65 720p videos/second and a latency of 5.7 seconds/video,true,"NVIDIA GB300 NVL72 (72 GPUs) WAN 2.2, Offline and SingleStream",0.65 videos/s and 5.7 s,Offline 0.6546 samples/s. SingleStream 5.808 s on GB300. The NVIDIA GB200 NVL72 row is 5.732 s,BELOW,Throughput matches. The GB300 latency is 5.8 s. The 5.7 s figure matches the GB200 row.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, NVIDIA statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+NV1,NVIDIA,Supp NVIDIA v6.1,"Vera Rubin NVL72 delivered up to 2.5x higher token throughput compared to GB300 NVL72 on DeepSeek-R1 across offline, server, and interactive scenarios",true,"NVIDIA Vera Rubin NVL72 (72 GPUs) against NVIDIA GB300 NVL72 (72 GPUs), DeepSeek-R1",up to 2.5x,"Offline 1.74x, Server 1.97x, Interactive 2.57x",CONSISTENT,The Vera Rubin entries have availability preview in the file.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, NVIDIA statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+NV2,NVIDIA,Supp NVIDIA v6.1,Vera Rubin NVL72 achieved up to 3.7x higher throughput compared to GB300 NVL72,true,"same two systems, Qwen3-VL-235B",up to 3.7x,"Offline 1.83x, Server 1.92x, Interactive 3.74x",CONSISTENT,The Vera Rubin entries have availability preview in the file.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, NVIDIA statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+NV3,NVIDIA,Supp NVIDIA v6.1,DeepSeek-R1 submission scaled to 288 GPUs across four GB300 NVL72 racks with near-linear scaling efficiency of 99% relative to the 72-GPU single-rack baseline for the offline scenario,true,NVIDIA GB300 NVL72 288-GPU DeepSeek-R1 Offline against 4 times the 72-GPU entry,99%,"2,705,130 against 4 x 679,740 = 99.5%",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, NVIDIA statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+NV6,NVIDIA,Supp NVIDIA v6.1,Qwen3-VL performance improved by up to 1.6x on GB300 systems,true,"NVIDIA GB300 Qwen3-VL-235B Server per GPU, v6.1 72-GPU entry against v6.0 entry 6.0-0077 (4 GPUs)",up to 1.6x,16.81 per GPU against 10.41 (1.62x),CONSISTENT,"Per GPU, Server. Offline per GPU is lower.",false,"MLCommons: MLPerf Inference v6.1 Results Discussion, NVIDIA statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+NV5,NVIDIA,Supp NVIDIA v6.1,9x higher throughput and 7.5x lower latency than single-node,true,NVIDIA GB300 NVL72 (72 GPUs) WAN 2.2 against the smaller entries in the file: Oracle GB300 4-GPU and NVIDIA DGX B300 8-GPU,"9x throughput, 7.5x latency","Against the 4-GPU GB300 entry: 47.3x throughput, 5.3x lower latency. Against the 8-GPU DGX B300 entry: 10.5x and 3.1x",CANNOT_TELL,Neither smaller entry gives 9x and 7.5x. The single-node baseline the statement uses is not in the file.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, NVIDIA statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+LA1,Lambda,Supp Lambda v6.1,GPT-OSS-120b particularly providing up to an 8.85% increase in throughput over the same workload during MLPerf Inference v6.0,true,"Lambda GB300 4-GPU entry, v6.0 entry 6.0-0063 against the v6.1 entry, GPT-OSS-120B",up to +8.85%,"Server 53,462.7 to 58,195.8 (+8.85%). Offline 60,220.4 to 65,511.9 (+8.79%)",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Lambda statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+SX1,ScitiX,Supp ScitiX v6.1,"the measured throughput reached 60,411 tokens/s in the Offline scenario and 59,668 tokens/s in the Server scenario",true,ScitiX B200 8-GPU DeepSeek-R1 entry,"60,411 and 59,668","60,410.9 and 59,668.3",CONSISTENT,The entries are in the open division.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, ScitiX statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+VB1,VibeHPC,Supp VibeHPC v6.1,"Our eight-GPU B300 system delivered 53,267.2 samples/second on Whisper Offline and 0.0783824 samples/second on WAN 2.2 Offline",true,VibeHPC 8-GPU B300 entry,"53,267.2 and 0.0783824","53,267.2 and 0.0783824",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, VibeHPC statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+VB2,VibeHPC,Supp VibeHPC v6.1,"the same system achieved 156,867 tokens/second in Server and 161,767 tokens/second in Offline",true,"same entry, Llama 3.1-8B","156,867 and 161,767","156,867 and 161,767",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, VibeHPC statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+VB3,VibeHPC,Supp VibeHPC v6.1,"achieving 20,420.4 tokens/second in Server and 21,266.1 tokens/second in Offline for Llama 3.1-8B",true,"VibeHPC single B300 entry, Llama 3.1-8B","20,420.4 and 21,266.1","20,420.4 and 21,266.1",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, VibeHPC statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+VB4,VibeHPC,Supp VibeHPC v6.1,"14,495.6 and 14,636.8 tokens/second for Llama 2-70B (99.9%), and 13,854.0 and 14,130.6 tokens/second for GPT-OSS-120B in Server and Offline, respectively",true,"same single B300 entry, Llama 2 70B 99.9 and GPT-OSS-120B","14,495.6 and 14,636.8, 13,854.0 and 14,130.6","Llama 2 70B 14,495.6 and 14,636.8. GPT-OSS-120B 13,854.0 and 14,130.6",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, VibeHPC statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+WI1,Wiwynn,Supp Wiwynn v6.1,"This node reached 63,991 tokens/second in the Offline scenario and 60,466 tokens/second in the Server scenario",true,"Wiwynn OneW GB300 4-GPU entry, GPT-OSS-120B","63,991 and 60,466","63,990.5 and 60,466.2",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Wiwynn statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+WI2,Wiwynn,Supp Wiwynn v6.1,"p99 TTFT was 336 ms against a 3,000 ms limit, with accuracy above 83% across both scenarios",true,accuracy fields of the Wiwynn entries,"p99 TTFT 336 ms, accuracy above 83%",Accuracy fields: exact_match: 83.440. exact_match: 83.187. The file has no time to first token,CANNOT_TELL,The accuracy figures are above 83. The time to first token is not in the file.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Wiwynn statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+OR1,Orrick Industries,Supp Orrick Industries v6.1,"274,927 output tokens per second",true,"Orrick entries in the v6.1 file, Llama 3.1-8B Offline, open division","274,927","274,927, six accelerators, availability rdi",CONSISTENT,The entry is in the open division with availability rdi.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Orrick Industries statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+OR2,Orrick Industries,Supp Orrick Industries v6.1,"130,928 output tokens per second",true,"same, Llama 2 70B 99.9 Offline","130,928","130,928",CONSISTENT,The entry is in the open division with availability rdi.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Orrick Industries statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+NK1,Individual contributor Naeem Khoshnevis,Supp Individual contributor v6.1,"The system achieved 7,796.59 tokens/s in the Offline scenario and 7,376.60 tokens/s in the Server scenario.",true,"Naeem Khoshnevis single H200 entry, Llama 3.1-8B","7,796.59 and 7,376.60","7,796.59 and 7,376.60",CONSISTENT,Availability in the file is rdi.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Individual contributor statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+NE1,Nebius,Supp Nebius v6.1,"The Nebius VR200 NVL72 preview ran on 36 GPUs across nine nodes. Nebius GB300 NVL72 submissions spanned an 18-node full 72 GPU rack, a two-node 8 GPU configuration, and a single-node 4 GPU configuration.",true,Nebius VR200 and GB300 entries in the v6.1 file,"36 GPUs on nine nodes, 72 GPUs on 18 nodes, 8 GPUs on two nodes, 4 GPUs on one node","Nebius GB300 NVL72 (4x GB300-288GB_aarch64, TensorRT) 4 GPUs 1 nodes available. Nebius GB300 NVL72 (72x GB300-288GB_aarch64, TensorRT) 72 GPUs 18 nodes available. Nebius GB300 NVL72 (8x GB300-288GB_aarch64, TensorRT) 8 GPUs 2 nodes available. Nebius VR200 NVL72 (36x VR200-288GB_aarch64, TensorRT) 36 GPUs 9 nodes preview",CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Nebius statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+NE2,Nebius,Supp Nebius v6.1,Nebius was one of two submitters with next-generation silicon.,true,organizations with preview entries in the v6.1 file,one of two,"Organizations with preview entries: NVIDIA, Nebius",CONSISTENT,The file marks the Vera Rubin entries as preview.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Nebius statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+OC1,Oracle,Supp Oracle v6.1,"Oracle's GB300 NVL72 submissions spanned a full 18-node, 72-GPU rack; a two-node, 8-GPU configuration; and a single-node, 4-GPU configuration.",true,Oracle GB300 entries in the v6.1 file,"72 GPUs on 18 nodes, 8 on two, 4 on one",4 GPUs on 1 nodes. 8 GPUs on 2 nodes. 72 GPUs on 18 nodes,CONSISTENT,,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, Oracle statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+GG1,GigaComputing,Supp GigaComputing v6.1,"GIGABYTE G894-ZD3-AAX7, using: 2x AMD EPYC 9575F (64 core) processors",true,"GigaComputing G894 entries in the v6.1 file, host processor fields","G894-ZD3 with 2 x EPYC 9575F (64 core), G894-AD3 with 2 x Xeon 6979P (120 core)","G894-AD3-AAX7: Intel(R) Xeon(R) 6979P, 2 x 56 cores. G894-ZD3-AAX7: AMD EPYC 9575F, 2 x 64 cores",BELOW,The processor models and the 64-core figure match. The file field for the 6979P system reads 56 cores per processor. The statement says 120.,false,"MLCommons: MLPerf Inference v6.1 Results Discussion, GigaComputing statement (supplement posted 16 September 2026)",https://mlcommons.org/wp-content/uploads/2026/09/Supplemental-MLCommons-MLPerf-Inference-v6.1.pdf,2026-10-28,MLPerf Inference v6.1
+M1,Apple,Apple Mac,614GB/s memory bandwidth,true,"Metal compute kernels on 1 GiB buffers (copy and triad, 3 runs of 20 passes) and a CPU triad on 1 GiB arrays (6, 12 and 18 threads, 3 runs of 15 passes), bandwidth counted as bytes read plus written over time",614 GB/s,"GPU triad 560.2 GB/s (91.2%), GPU copy 557.9 GB/s (90.9%), median GPU triad 559.4 GB/s. CPU only 362.9 GB/s (59.1%)",BELOW,"The page states a figure and no test method. The Mac was idle apart from this terminal, our other jobs were paused for the run, and the power mode was 2 (2 is High Power).",false,"Apple: MacBook Pro Technical Specifications, 14-inch with M5 Max",https://www.apple.com/macbook-pro/specs/,2026-10-28,"MacBook Pro 14-inch with M5 Max, measured on our unit"
+M2,Apple,Apple Mac,18-core CPU with: 6 super cores 12 performance cores,true,system_profiler and sysctl on the unit,"18 cores, 6 super and 12 performance",18 (6 Super and 12 Performance). sysctl reports 6 Super and 12 Performance,CONSISTENT,Counts are what the operating system reports.,false,"Apple: MacBook Pro Technical Specifications, 14-inch with M5 Max",https://www.apple.com/macbook-pro/specs/,2026-10-28,"MacBook Pro 14-inch with M5 Max, measured on our unit"
+M3,Apple,Apple Mac,40-core GPU,true,system_profiler on the unit,40-core GPU,40 cores,CONSISTENT,Counts are what the operating system reports.,false,"Apple: MacBook Pro Technical Specifications, 14-inch with M5 Max",https://www.apple.com/macbook-pro/specs/,2026-10-28,"MacBook Pro 14-inch with M5 Max, measured on our unit"
+M4,Apple,Apple Mac,"48GB, 64GB, or 128GB unified memory",true,sysctl hw.memsize on the unit,48GB,"51,539,607,552 bytes, 48 GiB",CONSISTENT,The unit has 48 GB. The 40-core GPU configuration is the one the page lists with 614 GB/s.,false,"Apple: MacBook Pro Technical Specifications, 14-inch with M5 Max",https://www.apple.com/macbook-pro/specs/,2026-10-28,"MacBook Pro 14-inch with M5 Max, measured on our unit"
+M5,Apple,Apple Mac,"M5 Max 2TB, 4TB, or 8TB SSD",true,system_profiler on the unit,4TB,"3,996,276,899,840 bytes, 4.00 TB",CONSISTENT,Capacity of the data volume as the operating system reports it.,false,"Apple: MacBook Pro Technical Specifications, 14-inch with M5 Max",https://www.apple.com/macbook-pro/specs/,2026-10-28,"MacBook Pro 14-inch with M5 Max, measured on our unit"
+M6,Apple,Apple Mac,3024-by-1964 native resolution at 254 pixels per inch,true,system_profiler on the unit,3024 by 1964,3024 x 1964,CONSISTENT,Resolution only. Pixels per inch was not checked.,false,"Apple: MacBook Pro Technical Specifications, 14-inch with M5 Max",https://www.apple.com/macbook-pro/specs/,2026-10-28,"MacBook Pro 14-inch with M5 Max, measured on our unit"
+M7,Apple,Apple Mac,72.4-watt-hour lithium-polymer battery,true,battery gauge through ioreg on the unit,72.4 Wh,"gauge reports design capacity 6,249 mAh and full-charge capacity 6,389 mAh (102.2% of design) after 84 cycles",CANNOT_TELL,"The gauge reports milliamp-hours. Watt-hours needs the pack voltage, which the gauge does not report as a rating.",false,"Apple: MacBook Pro Technical Specifications, 14-inch with M5 Max",https://www.apple.com/macbook-pro/specs/,2026-10-28,"MacBook Pro 14-inch with M5 Max, measured on our unit"
+M8,Apple,Apple Mac,Up to 20 hours video streaming,true,a full discharge under Apple's stated test conditions,up to 20 hours,not measured,CANNOT_TELL,Apple tested a preproduction system with set display brightness. The unit here was not run down.,false,"Apple: MacBook Pro Technical Specifications, 14-inch with M5 Max",https://www.apple.com/macbook-pro/specs/,2026-10-28,"MacBook Pro 14-inch with M5 Max, measured on our unit"
+M9,Apple,Apple Mac,XDR brightness: up to 1000 nits sustained full-screen,true,a luminance meter on the display,up to 1000 nits,not measured,CANNOT_TELL,No luminance meter was used.,false,"Apple: MacBook Pro Technical Specifications, 14-inch with M5 Max",https://www.apple.com/macbook-pro/specs/,2026-10-28,"MacBook Pro 14-inch with M5 Max, measured on our unit"