diff --git "a/CATALOG.json" "b/CATALOG.json" --- "a/CATALOG.json" +++ "b/CATALOG.json" @@ -2,6 +2,8 @@ { "name": "activation-recompute-fusion", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 2356.1, "gpus": 1, @@ -23,6 +25,8 @@ { "name": "adafactor-factored-update", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 520.8, "gpus": 1, @@ -42,6 +46,8 @@ { "name": "adaptive-sparsity-threshold", "family": "Video \u2014 sparse / efficient attention", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 359.1, "gpus": 1, @@ -62,6 +68,8 @@ { "name": "alibi-slopes-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -78,6 +86,8 @@ { "name": "altup-predict-step", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "GB/s", "roofline_us": 651.0, "gpus": 1, @@ -96,6 +106,8 @@ { "name": "any-res-image-split", "family": "Multimodal & audio", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 336.5, "gpus": 1, @@ -114,6 +126,8 @@ { "name": "async-weight-prefetch-pipeline", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -134,6 +148,8 @@ { "name": "attention-lse-merge", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "GB/s", "roofline_us": 1057.9, "gpus": 1, @@ -153,6 +169,8 @@ { "name": "attention-qk-norm-inline", "family": "Normalization, RoPE & elementwise fusion", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -172,6 +190,8 @@ { "name": "attention-sink-decode", "family": "KV cache & paging", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "GB/s", "roofline_us": 416.7, "gpus": 1, @@ -192,6 +212,8 @@ { "name": "attn-density-profile-fused", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 1057.0, "gpus": 1, @@ -212,6 +234,8 @@ { "name": "audio-codec-rvq-quantize", "family": "Multimodal & audio", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 431.4, "gpus": 1, @@ -231,6 +255,8 @@ { "name": "audio-mel-spectrogram-fused", "family": "Multimodal & audio", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "GB/s", "roofline_us": 315.9, "gpus": 1, @@ -249,6 +275,8 @@ { "name": "awq-dequant-gemm", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -267,6 +295,8 @@ { "name": "based-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 533.2, "gpus": 1, @@ -284,6 +314,8 @@ { "name": "based-forward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 480.1, "gpus": 1, @@ -301,6 +333,8 @@ { "name": "beam-search-step", "family": "Sampling & speculative decoding", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 416.7, "gpus": 1, @@ -319,6 +353,8 @@ { "name": "block-sparse-video-attn", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 1346.6, "gpus": 1, @@ -339,6 +375,8 @@ { "name": "blockwise-fp8-gemm", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -359,6 +397,8 @@ { "name": "cache-hit-skip-gate", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 593.4, "gpus": 1, @@ -376,6 +416,8 @@ { "name": "cascade-prefix-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -395,6 +437,8 @@ { "name": "causal-conv1d-decode-step", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 625.0, "gpus": 1, @@ -412,6 +456,8 @@ { "name": "causal-conv1d-fused", "family": "Linear attention & SSM", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 833.3, "gpus": 1, @@ -429,6 +475,8 @@ { "name": "causal-conv3d-backward", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 728.0, "gpus": 1, @@ -450,6 +498,8 @@ { "name": "causal-conv3d-cache-step", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 465.9, "gpus": 1, @@ -471,6 +521,8 @@ { "name": "causal-conv3d-forward", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 533.9, "gpus": 1, @@ -491,6 +543,8 @@ { "name": "causal-conv3d-fp8", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 640.6, "gpus": 1, @@ -512,6 +566,8 @@ { "name": "causal-video-kv-cache-decode", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 2503.4, "gpus": 1, @@ -532,6 +588,8 @@ { "name": "cfg-distill-single-pass", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 494.5, "gpus": 1, @@ -550,6 +608,8 @@ { "name": "chunked-prefill-attention", "family": "KV cache & paging", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 520.8, "gpus": 1, @@ -568,6 +628,8 @@ { "name": "chunked-prefill-gather", "family": "Attention \u2014 text LLM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 472.2, "gpus": 1, @@ -586,6 +648,8 @@ { "name": "chunked-zloss-cross-entropy", "family": "Training, optimizer & RL", + "tier": "T3", + "tier_why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", "metric": "TFLOP/s", "roofline_us": 113092.6, "gpus": 1, @@ -603,6 +667,8 @@ { "name": "clip-logit-scale-fused", "family": "Multimodal & audio", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -621,6 +687,8 @@ { "name": "cogvideox-expert-adaln-block", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "GB/s", "roofline_us": 419.8, "gpus": 1, @@ -641,6 +709,8 @@ { "name": "cogvideox-joint-3d-attention", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 5546.9, "gpus": 1, @@ -662,6 +732,8 @@ { "name": "cogvideox-shared-ffn-expert-gate", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 3883.2, "gpus": 1, @@ -683,6 +755,8 @@ { "name": "cogvideox-sincos-3d-embed", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 419.8, "gpus": 1, @@ -703,6 +777,8 @@ { "name": "cogvideox-vae-spatial-norm3d", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "GB/s", "roofline_us": 511.8, "gpus": 1, @@ -724,6 +800,8 @@ { "name": "comba-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 441.8, "gpus": 1, @@ -741,6 +819,8 @@ { "name": "comba-decode-step", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 473.3, "gpus": 1, @@ -759,6 +839,8 @@ { "name": "comba-forward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 220.9, "gpus": 1, @@ -775,6 +857,8 @@ { "name": "compressed-kv-video-attn", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 1187.3, "gpus": 1, @@ -795,6 +879,8 @@ { "name": "conformer-conv-module", "family": "Multimodal & audio", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 871.5, "gpus": 1, @@ -815,6 +901,8 @@ { "name": "conv3d-layout-transform", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 358.9, "gpus": 1, @@ -835,6 +923,8 @@ { "name": "conv3d-tiled-decode-blend", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 943.5, "gpus": 1, @@ -857,6 +947,8 @@ { "name": "cross-attention-varlen", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", "metric": "TFLOP/s", "roofline_us": 571.1, "gpus": 1, @@ -877,6 +969,8 @@ { "name": "cross-layer-fusion-2layer", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -896,6 +990,8 @@ { "name": "ddim-inversion-step", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 593.4, "gpus": 1, @@ -914,6 +1010,8 @@ { "name": "deepseek-mla-lowrank-proj", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 1237.9, "gpus": 1, @@ -935,6 +1033,8 @@ { "name": "deepseek-mla-proj-backward", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 2830.1, "gpus": 1, @@ -956,6 +1056,8 @@ { "name": "deepseek-mla-vabsorb-outproj", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "MLA V-absorb folds the value projection into the output projection: a fused GEMM pair, not a plain one", "metric": "TFLOP/s", "roofline_us": 736.3, "gpus": 1, @@ -976,6 +1078,8 @@ { "name": "deepseek-node-dispatch-pack", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 456.0, "gpus": 1, @@ -996,6 +1100,8 @@ { "name": "deepseek-v32-indexer-decode", "family": "KV cache & paging", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "GB/s", "roofline_us": 314.6, "gpus": 1, @@ -1016,6 +1122,8 @@ { "name": "deepseek-v32-sparse-mla-decode", "family": "KV cache & paging", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "GB/s", "roofline_us": 570.7, "gpus": 1, @@ -1037,6 +1145,8 @@ { "name": "delta-product-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 398.8, "gpus": 1, @@ -1054,6 +1164,8 @@ { "name": "delta-rule-forward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 539.9, "gpus": 1, @@ -1071,6 +1183,8 @@ { "name": "deltaformer-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 1570.7, "gpus": 1, @@ -1088,6 +1202,8 @@ { "name": "deltaformer-forward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 3141.5, "gpus": 1, @@ -1104,6 +1220,8 @@ { "name": "depthwise-separable-conv3d", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 396.3, "gpus": 1, @@ -1125,6 +1243,8 @@ { "name": "diff-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 1570.7, "gpus": 1, @@ -1141,6 +1261,8 @@ { "name": "dist-allgather-gemm-overlap", "family": "Distributed \u2014 multi-GPU collectives", + "tier": "T3", + "tier_why": "2-GPU compute/communication overlap: the GEMM is a library call but the overlap schedule is the task", "metric": "TFLOP/s", "roofline_us": 1570.7, "gpus": 2, @@ -1161,6 +1283,8 @@ { "name": "dist-allreduce-rmsnorm", "family": "Distributed \u2014 multi-GPU collectives", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 14062.5, "gpus": 2, @@ -1180,6 +1304,8 @@ { "name": "dist-moe-a2a-dispatch", "family": "Distributed \u2014 multi-GPU collectives", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 4687.5, "gpus": 2, @@ -1200,6 +1326,8 @@ { "name": "dist-oneshot-allreduce", "family": "Distributed \u2014 multi-GPU collectives", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 3750.0, "gpus": 2, @@ -1220,6 +1348,8 @@ { "name": "dist-reduce-scatter-rmsnorm", "family": "Distributed \u2014 multi-GPU collectives", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 8437.5, "gpus": 2, @@ -1240,6 +1370,8 @@ { "name": "dist-tp-embedding-allreduce", "family": "Distributed \u2014 multi-GPU collectives", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 14062.5, "gpus": 2, @@ -1260,6 +1392,8 @@ { "name": "distill-kl-logits-fused", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 1223.1, "gpus": 1, @@ -1279,6 +1413,8 @@ { "name": "dora-magnitude-decompose", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 813.2, "gpus": 1, @@ -1298,6 +1434,8 @@ { "name": "dpmsolverpp-step", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 494.5, "gpus": 1, @@ -1316,6 +1454,8 @@ { "name": "dpo-loss-fused", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 1449.0, "gpus": 1, @@ -1334,6 +1474,8 @@ { "name": "draft-tree-build", "family": "Sampling & speculative decoding", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 478.6, "gpus": 1, @@ -1352,6 +1494,8 @@ { "name": "dsa-sparse-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 294.5, "gpus": 1, @@ -1368,6 +1512,8 @@ { "name": "dyt-dynamic-tanh-backward", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 703.1, "gpus": 1, @@ -1386,6 +1532,8 @@ { "name": "dyt-dynamic-tanh-norm", "family": "Normalization, RoPE & elementwise fusion", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 468.8, "gpus": 1, @@ -1404,6 +1552,8 @@ { "name": "embedding-backward-scatter", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 1041.9, "gpus": 1, @@ -1422,6 +1572,8 @@ { "name": "entropy-bonus-reduce", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 1223.1, "gpus": 1, @@ -1442,6 +1594,8 @@ { "name": "feature-cache-blend", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 741.7, "gpus": 1, @@ -1459,6 +1613,8 @@ { "name": "flash-attn-backward", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -1481,6 +1637,8 @@ { "name": "flex-blockmask-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", "metric": "TFLOP/s", "roofline_us": 583.1, "gpus": 1, @@ -1499,6 +1657,8 @@ { "name": "flow-match-euler-step", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 494.5, "gpus": 1, @@ -1516,6 +1676,8 @@ { "name": "flow-match-loss-backward", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 544.1, "gpus": 1, @@ -1536,6 +1698,8 @@ { "name": "flux-attention-backward", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 7190.2, "gpus": 1, @@ -1555,6 +1719,8 @@ { "name": "flux-attn-fp8", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 5752.2, "gpus": 1, @@ -1575,6 +1741,8 @@ { "name": "flux-double-stream-block", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 10478.2, "gpus": 1, @@ -1598,6 +1766,8 @@ { "name": "flux-dual-stream-mlp", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 5522.1, "gpus": 1, @@ -1619,6 +1789,8 @@ { "name": "flux-final-layer-unpatchify", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "GB/s", "roofline_us": 478.5, "gpus": 1, @@ -1640,6 +1812,8 @@ { "name": "flux-fused-qkv-mlp-proj", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 4831.8, "gpus": 1, @@ -1661,6 +1835,8 @@ { "name": "flux-joint-attention", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 5752.2, "gpus": 1, @@ -1680,6 +1856,8 @@ { "name": "flux-modulation-backward", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "GB/s", "roofline_us": 395.5, "gpus": 1, @@ -1702,6 +1880,8 @@ { "name": "flux-modulation-fused", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "GB/s", "roofline_us": 366.2, "gpus": 1, @@ -1723,6 +1903,8 @@ { "name": "flux-qk-rmsnorm-rope", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 395.5, "gpus": 1, @@ -1744,6 +1926,8 @@ { "name": "flux-rope2d-apply", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 366.2, "gpus": 1, @@ -1763,6 +1947,8 @@ { "name": "flux-single-stream-block", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 10478.2, "gpus": 1, @@ -1786,6 +1972,8 @@ { "name": "forgetting-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -1802,6 +1990,8 @@ { "name": "forgetting-attn-backward", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 883.5, "gpus": 1, @@ -1819,6 +2009,8 @@ { "name": "fp4-stochastic-round", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 413.8, "gpus": 1, @@ -1838,6 +2030,8 @@ { "name": "fp8-attention-backward", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 1963.4, "gpus": 1, @@ -1856,6 +2050,8 @@ { "name": "fp8-blockwise-gemm", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -1873,6 +2069,8 @@ { "name": "fp8-flash-attention", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -1889,6 +2087,8 @@ { "name": "fp8-grouped-gemm-masked", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 687.2, "gpus": 1, @@ -1908,6 +2108,8 @@ { "name": "fp8-kv-decode-attention", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "GB/s", "roofline_us": 214.8, "gpus": 1, @@ -1926,6 +2128,8 @@ { "name": "fp8-per-token-quant-transpose", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 423.2, "gpus": 1, @@ -1945,6 +2149,8 @@ { "name": "fp8-quant-dequant-fused", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 366.2, "gpus": 1, @@ -1964,6 +2170,8 @@ { "name": "frame-anchor-attention", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 3779.6, "gpus": 1, @@ -1983,6 +2191,8 @@ { "name": "fsdp-allgather-unflatten", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 400.1, "gpus": 1, @@ -2004,6 +2214,8 @@ { "name": "fsdp-flatten-shard-params", "family": "Training, optimizer & RL", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 585.2, "gpus": 1, @@ -2024,6 +2236,8 @@ { "name": "fsdp-mixed-precision-cast", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 453.2, "gpus": 1, @@ -2046,6 +2260,8 @@ { "name": "fsdp-reduce-scatter-bucket", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 568.8, "gpus": 1, @@ -2068,6 +2284,8 @@ { "name": "fused-adamw-multitensor", "family": "Training, optimizer & RL", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 4050.7, "gpus": 1, @@ -2086,6 +2304,8 @@ { "name": "fused-linear-cross-entropy", "family": "Training, optimizer & RL", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 221325.8, "gpus": 1, @@ -2102,6 +2322,8 @@ { "name": "fused-muon-newton-schulz-batched", "family": "Training, optimizer & RL", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 6626.5, "gpus": 1, @@ -2120,6 +2342,8 @@ { "name": "fused-qk-norm-rope", "family": "Normalization, RoPE & elementwise fusion", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 361.3, "gpus": 1, @@ -2138,6 +2362,8 @@ { "name": "fused-residual-rmsnorm", "family": "Normalization, RoPE & elementwise fusion", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 312.5, "gpus": 1, @@ -2155,6 +2381,8 @@ { "name": "fused-rmsnorm-fp8-quant", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 473.6, "gpus": 1, @@ -2172,6 +2400,8 @@ { "name": "fused-topk-topp-filter", "family": "Sampling & speculative decoding", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 416.7, "gpus": 1, @@ -2192,6 +2422,8 @@ { "name": "gae-advantage-scan", "family": "Training, optimizer & RL", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 546.9, "gpus": 1, @@ -2211,6 +2443,8 @@ { "name": "gated-delta-product-forward", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "TFLOP/s", "roofline_us": 478.6, "gpus": 1, @@ -2227,6 +2461,8 @@ { "name": "gated-deltanet-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -2247,6 +2483,8 @@ { "name": "gated-deltanet-decode-step", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 472.4, "gpus": 1, @@ -2265,6 +2503,8 @@ { "name": "gated-deltanet-forward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -2283,6 +2523,8 @@ { "name": "geglu-backward-fused", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 781.3, "gpus": 1, @@ -2302,6 +2544,8 @@ { "name": "gla-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 441.8, "gpus": 1, @@ -2319,6 +2563,8 @@ { "name": "gla-decode-step", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 474.2, "gpus": 1, @@ -2337,6 +2583,8 @@ { "name": "gla-forward", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "TFLOP/s", "roofline_us": 441.8, "gpus": 1, @@ -2353,6 +2601,8 @@ { "name": "gptq-actorder-gemm", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -2373,6 +2623,8 @@ { "name": "gqa-decode-attention", "family": "KV cache & paging", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "GB/s", "roofline_us": 833.3, "gpus": 1, @@ -2391,6 +2643,8 @@ { "name": "gqa-paged-decode-attention", "family": "KV cache & paging", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "GB/s", "roofline_us": 626.4, "gpus": 1, @@ -2411,6 +2665,8 @@ { "name": "grad-global-norm-clip", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 2842.4, "gpus": 1, @@ -2428,6 +2684,8 @@ { "name": "grad-topk-error-feedback", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 319.0, "gpus": 1, @@ -2447,6 +2705,8 @@ { "name": "gradient-accumulation-fused", "family": "Training, optimizer & RL", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 2095.5, "gpus": 1, @@ -2464,6 +2724,8 @@ { "name": "grammar-jump-ahead-compile", "family": "KV cache & paging", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 417.4, "gpus": 1, @@ -2485,6 +2747,8 @@ { "name": "gridwide-barrier-primitive", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -2503,6 +2767,8 @@ { "name": "grouped-fp8-gemm", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -2525,6 +2791,8 @@ { "name": "groupnorm3d-backward", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 418.7, "gpus": 1, @@ -2546,6 +2814,8 @@ { "name": "groupnorm3d-forward", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 370.3, "gpus": 1, @@ -2566,6 +2836,8 @@ { "name": "grpo-advantage-whiten", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 325.5, "gpus": 1, @@ -2584,6 +2856,8 @@ { "name": "grpo-logprob-ratio", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 1223.1, "gpus": 1, @@ -2603,6 +2877,8 @@ { "name": "gsa-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 441.8, "gpus": 1, @@ -2620,6 +2896,8 @@ { "name": "gsa-forward", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "TFLOP/s", "roofline_us": 392.7, "gpus": 1, @@ -2636,6 +2914,8 @@ { "name": "guided-decoding-mask", "family": "Sampling & speculative decoding", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 317.4, "gpus": 1, @@ -2656,6 +2936,8 @@ { "name": "hgrn2-forward", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "TFLOP/s", "roofline_us": 441.8, "gpus": 1, @@ -2672,6 +2954,8 @@ { "name": "hierarchical-coarse-fine-attn", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 2279.0, "gpus": 1, @@ -2692,6 +2976,8 @@ { "name": "hifigan-mrf-resblock", "family": "Multimodal & audio", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 517.7, "gpus": 1, @@ -2712,6 +2998,8 @@ { "name": "hunyuan-attention-backward", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 9498.2, "gpus": 1, @@ -2732,6 +3020,8 @@ { "name": "hunyuan-attn-fp8", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 10602.4, "gpus": 1, @@ -2752,6 +3042,8 @@ { "name": "hunyuan-dualstream-attn-proj", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T3", + "tier_why": "two per-stream output projections plus per-sample gates: a 2-group GEMM with M=1e5 against M=1e2", "metric": "TFLOP/s", "roofline_us": 2222.6, "gpus": 1, @@ -2773,6 +3065,8 @@ { "name": "hunyuan-dualstream-block", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 10013.4, "gpus": 1, @@ -2796,6 +3090,8 @@ { "name": "hunyuan-final-layer-unpatchify", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "GB/s", "roofline_us": 433.7, "gpus": 1, @@ -2816,6 +3112,8 @@ { "name": "hunyuan-joint-attention", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 10657.7, "gpus": 1, @@ -2836,6 +3134,8 @@ { "name": "hunyuan-mmdit-modulation", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "GB/s", "roofline_us": 284.5, "gpus": 1, @@ -2856,6 +3156,8 @@ { "name": "hunyuan-modulation-backward", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 585.9, "gpus": 1, @@ -2878,6 +3180,8 @@ { "name": "hunyuan-qk-norm-fused", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 853.4, "gpus": 1, @@ -2899,6 +3203,8 @@ { "name": "hunyuan-rope3d-apply", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 643.1, "gpus": 1, @@ -2918,6 +3224,8 @@ { "name": "hunyuan-singlestream-block", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 10013.4, "gpus": 1, @@ -2941,6 +3249,8 @@ { "name": "hunyuan-split-ffn", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 9939.8, "gpus": 1, @@ -2962,6 +3272,8 @@ { "name": "hunyuan-split-ffn-backward", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 7951.8, "gpus": 1, @@ -2983,6 +3295,8 @@ { "name": "hunyuan-token-replace-i2v", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "GB/s", "roofline_us": 566.5, "gpus": 1, @@ -3004,6 +3318,8 @@ { "name": "hunyuan-varlen-multiresolution", "family": "Video \u2014 HunyuanVideo MMDiT", + "tier": "T3", + "tier_why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", "metric": "TFLOP/s", "roofline_us": 9484.7, "gpus": 1, @@ -3025,6 +3341,8 @@ { "name": "hybrid-layer-output-gate", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 1767.1, "gpus": 1, @@ -3045,6 +3363,8 @@ { "name": "hybrid-ssm-attn-parallel-decode", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "GB/s", "roofline_us": 731.4, "gpus": 1, @@ -3065,6 +3385,8 @@ { "name": "image-token-merge-pixelunshuffle", "family": "Multimodal & audio", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 421.1, "gpus": 1, @@ -3083,6 +3405,8 @@ { "name": "instruction-interpreter-dispatch", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3102,6 +3426,8 @@ { "name": "int4-kv-cache-quant", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 546.9, "gpus": 1, @@ -3120,6 +3446,8 @@ { "name": "int4-weight-only-decode-gemv", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "GB/s", "roofline_us": 218.2, "gpus": 1, @@ -3139,6 +3467,8 @@ { "name": "int8-w8a8-gemm", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -3157,6 +3487,8 @@ { "name": "istft-overlap-add", "family": "Multimodal & audio", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 516.7, "gpus": 1, @@ -3176,6 +3508,8 @@ { "name": "kda-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 539.9, "gpus": 1, @@ -3193,6 +3527,8 @@ { "name": "kda-decode-step", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 474.3, "gpus": 1, @@ -3212,6 +3548,8 @@ { "name": "kda-forward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 539.9, "gpus": 1, @@ -3229,6 +3567,8 @@ { "name": "kv-block-pack-for-transfer", "family": "KV cache & paging", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 586.5, "gpus": 1, @@ -3249,6 +3589,8 @@ { "name": "kv-cache-4bit-groupwise", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "GB/s", "roofline_us": 520.8, "gpus": 1, @@ -3269,6 +3611,8 @@ { "name": "kv-cache-block-compaction", "family": "KV cache & paging", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 625.0, "gpus": 1, @@ -3288,6 +3632,8 @@ { "name": "kv-cache-fp8-quant-paged", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 473.6, "gpus": 1, @@ -3307,6 +3653,8 @@ { "name": "kv-cache-page-alloc", "family": "KV cache & paging", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 417.9, "gpus": 1, @@ -3326,6 +3674,8 @@ { "name": "kv-layout-contig-to-paged", "family": "KV cache & paging", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 636.2, "gpus": 1, @@ -3346,6 +3696,8 @@ { "name": "kv-repage-block-size-convert", "family": "KV cache & paging", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 645.3, "gpus": 1, @@ -3366,6 +3718,8 @@ { "name": "latent-blend-i2v-mask", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 509.9, "gpus": 1, @@ -3384,6 +3738,8 @@ { "name": "latent-interpolate-keyframe", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 314.7, "gpus": 1, @@ -3402,6 +3758,8 @@ { "name": "latent-normalize-perchannel", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 385.5, "gpus": 1, @@ -3423,6 +3781,8 @@ { "name": "latent-patchify-3d", "family": "Video \u2014 sparse / efficient attention", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 300.0, "gpus": 1, @@ -3441,6 +3801,8 @@ { "name": "laurel-lowrank-residual", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "GB/s", "roofline_us": 521.0, "gpus": 1, @@ -3460,6 +3822,8 @@ { "name": "layernorm-residual-fp8-backward", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 824.0, "gpus": 1, @@ -3478,6 +3842,8 @@ { "name": "layerscale-residual-fused", "family": "Normalization, RoPE & elementwise fusion", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 703.1, "gpus": 1, @@ -3496,6 +3862,8 @@ { "name": "lightning-attn-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 441.8, "gpus": 1, @@ -3512,6 +3880,8 @@ { "name": "lightning-attn-forward", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "TFLOP/s", "roofline_us": 343.6, "gpus": 1, @@ -3528,6 +3898,8 @@ { "name": "log-linear-attn-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 441.8, "gpus": 1, @@ -3545,6 +3917,8 @@ { "name": "log-linear-attn-forward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 441.8, "gpus": 1, @@ -3561,6 +3935,8 @@ { "name": "logits-gather-lastpos", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", "metric": "TFLOP/s", "roofline_us": 1570.7, "gpus": 1, @@ -3580,6 +3956,8 @@ { "name": "lora-backward-fused", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 822.2, "gpus": 1, @@ -3599,6 +3977,8 @@ { "name": "lora-fused-gemm", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 800.7, "gpus": 1, @@ -3617,6 +3997,8 @@ { "name": "ltx-fractional-rope3d", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 510.3, "gpus": 1, @@ -3638,6 +4020,8 @@ { "name": "ltx-highcompression-vae-decode", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 942.2, "gpus": 1, @@ -3659,6 +4043,8 @@ { "name": "ltx-per-token-adanorm", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 429.2, "gpus": 1, @@ -3679,6 +4065,8 @@ { "name": "ltx-per-token-adanorm-backward", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 333.8, "gpus": 1, @@ -3700,6 +4088,8 @@ { "name": "ltx-rectified-flow-shift", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 300.9, "gpus": 1, @@ -3720,6 +4110,8 @@ { "name": "mamba2-ssd-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 343.6, "gpus": 1, @@ -3737,6 +4129,8 @@ { "name": "mamba2-ssd-decode-step", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 474.3, "gpus": 1, @@ -3756,6 +4150,8 @@ { "name": "mamba2-ssd-forward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3775,6 +4171,8 @@ { "name": "megakernel-batch4-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3793,6 +4191,8 @@ { "name": "megakernel-gated-deltanet-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3814,6 +4214,8 @@ { "name": "megakernel-gqa-paged-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3833,6 +4235,8 @@ { "name": "megakernel-int4-weight-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3854,6 +4258,8 @@ { "name": "megakernel-llama1b-bf16-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3872,6 +4278,8 @@ { "name": "megakernel-llama1b-fp8-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3890,6 +4298,8 @@ { "name": "megakernel-llama1b-longctx-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3908,6 +4318,8 @@ { "name": "megakernel-llama8b-bf16-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3927,6 +4339,8 @@ { "name": "megakernel-lm-head-sample-inline", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3946,6 +4360,8 @@ { "name": "megakernel-mamba-hybrid-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3966,6 +4382,8 @@ { "name": "megakernel-mtp-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -3985,6 +4403,8 @@ { "name": "megakernel-nvfp4-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -4006,6 +4426,8 @@ { "name": "megakernel-prefill-chunk", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -4024,6 +4446,8 @@ { "name": "megakernel-qwen3-8b-fp8-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -4044,6 +4468,8 @@ { "name": "megakernel-qwen3-moe-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -4063,6 +4489,8 @@ { "name": "megakernel-spec-decode-verify", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -4082,6 +4510,8 @@ { "name": "megakernel-untied-head-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -4101,6 +4531,8 @@ { "name": "mesa-net-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 291.4, "gpus": 1, @@ -4118,6 +4550,8 @@ { "name": "mesa-net-forward", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "TFLOP/s", "roofline_us": 233.2, "gpus": 1, @@ -4134,6 +4568,8 @@ { "name": "min-p-sampling-filter", "family": "Sampling & speculative decoding", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 416.7, "gpus": 1, @@ -4153,6 +4589,8 @@ { "name": "mixed-precision-gemm-split", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 809.9, "gpus": 1, @@ -4173,6 +4611,8 @@ { "name": "mla-backward", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 1178.0, "gpus": 1, @@ -4190,6 +4630,8 @@ { "name": "mla-decode", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -4211,6 +4653,8 @@ { "name": "mla-decode-paged", "family": "KV cache & paging", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "GB/s", "roofline_us": 350.1, "gpus": 1, @@ -4232,6 +4676,8 @@ { "name": "mla-prefill", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 1030.8, "gpus": 1, @@ -4248,6 +4694,8 @@ { "name": "mm-embed-merge-scatter", "family": "Multimodal & audio", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 410.2, "gpus": 1, @@ -4266,6 +4714,8 @@ { "name": "moba-backward", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 540.4, "gpus": 1, @@ -4284,6 +4734,8 @@ { "name": "moba-block-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -4303,6 +4755,8 @@ { "name": "mochi-asymm-joint-attention", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T3", + "tier_why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", "metric": "TFLOP/s", "roofline_us": 5107.7, "gpus": 1, @@ -4324,6 +4778,8 @@ { "name": "mochi-asymm-qkv-proj", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 3621.9, "gpus": 1, @@ -4345,6 +4801,8 @@ { "name": "mochi-tanh-gated-rmsnorm", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 429.2, "gpus": 1, @@ -4366,6 +4824,8 @@ { "name": "mochi-visual-text-split-mlp", "family": "Video \u2014 CogVideoX / Mochi / LTX", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 4829.2, "gpus": 1, @@ -4388,6 +4848,8 @@ { "name": "moe-capacity-drop-dispatch", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 483.3, "gpus": 1, @@ -4406,6 +4868,8 @@ { "name": "moe-combine-backward", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 581.3, "gpus": 1, @@ -4423,6 +4887,8 @@ { "name": "moe-combine-weighted", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 205.2, "gpus": 1, @@ -4440,6 +4906,8 @@ { "name": "moe-expert-bias-update", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 260.5, "gpus": 1, @@ -4457,6 +4925,8 @@ { "name": "moe-expert-weight-gather", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 729.2, "gpus": 1, @@ -4475,6 +4945,8 @@ { "name": "moe-group-limited-routing", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 221.4, "gpus": 1, @@ -4492,6 +4964,8 @@ { "name": "moe-grouped-gemm-contiguous", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T3", + "tier_why": "grouped GEMM over a contiguous expert layout: no single library call covers it", "metric": "TFLOP/s", "roofline_us": 1374.4, "gpus": 1, @@ -4509,6 +4983,8 @@ { "name": "moe-grouped-gemm-varm", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T3", + "tier_why": "variable-M grouped GEMM: cuBLAS has no such call", "metric": "TFLOP/s", "roofline_us": 1570.7, "gpus": 1, @@ -4526,6 +5002,8 @@ { "name": "moe-grouped-swiglu", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T3", + "tier_why": "grouped GEMM with a fused SwiGLU epilogue", "metric": "TFLOP/s", "roofline_us": 1570.7, "gpus": 1, @@ -4543,6 +5021,8 @@ { "name": "moe-router-backward", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 429.5, "gpus": 1, @@ -4561,6 +5041,8 @@ { "name": "moe-shared-expert-fused", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 1032.1, "gpus": 1, @@ -4579,6 +5061,8 @@ { "name": "moe-token-permute", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 205.2, "gpus": 1, @@ -4596,6 +5080,8 @@ { "name": "moe-topk-softmax-fused", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 276.7, "gpus": 1, @@ -4613,6 +5099,8 @@ { "name": "mrope-multimodal-apply", "family": "Multimodal & audio", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 325.5, "gpus": 1, @@ -4631,6 +5119,8 @@ { "name": "mtp-head-forward", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 3337.8, "gpus": 1, @@ -4650,6 +5140,8 @@ { "name": "mtp-multi-head-loss", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 1666.7, "gpus": 1, @@ -4669,6 +5161,8 @@ { "name": "multi-lora-batched-decode", "family": "KV cache & paging", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 477.7, "gpus": 1, @@ -4689,6 +5183,8 @@ { "name": "muon-newton-schulz", "family": "Training, optimizer & RL", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -4709,6 +5205,8 @@ { "name": "mxfp4-dequant-gemm", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -4726,6 +5224,8 @@ { "name": "mxfp4-training-backward", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -4746,6 +5246,8 @@ { "name": "mxfp8-blockwise-gemm", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -4765,6 +5267,8 @@ { "name": "mxfp8-dual-axis-quantize", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 423.2, "gpus": 1, @@ -4785,6 +5289,8 @@ { "name": "ngpt-l2norm-residual", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 703.1, "gpus": 1, @@ -4803,6 +5309,8 @@ { "name": "ngram-lookahead-match", "family": "Sampling & speculative decoding", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 300.5, "gpus": 1, @@ -4821,6 +5329,8 @@ { "name": "noise-add-scaled", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 494.5, "gpus": 1, @@ -4839,6 +5349,8 @@ { "name": "nsa-compression-branch", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 393.1, "gpus": 1, @@ -4857,6 +5369,8 @@ { "name": "nsa-selected-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -4876,6 +5390,8 @@ { "name": "nsa-selected-backward", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 392.7, "gpus": 1, @@ -4893,6 +5409,8 @@ { "name": "nvfp4-dequant-gemm", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -4910,6 +5428,8 @@ { "name": "nvfp4-gemm-backward", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -4930,6 +5450,8 @@ { "name": "nvfp4-quantize", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 534.7, "gpus": 1, @@ -4948,6 +5470,8 @@ { "name": "optimizer-state-reshard", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 802.4, "gpus": 1, @@ -4970,6 +5494,8 @@ { "name": "paged-kv-append", "family": "KV cache & paging", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 629.9, "gpus": 1, @@ -4989,6 +5515,8 @@ { "name": "path-attn-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 865.1, "gpus": 1, @@ -5006,6 +5534,8 @@ { "name": "path-attn-forward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 805.3, "gpus": 1, @@ -5022,6 +5552,8 @@ { "name": "penalty-count-bincount-build", "family": "Sampling & speculative decoding", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 417.9, "gpus": 1, @@ -5041,6 +5573,8 @@ { "name": "persistent-kv-append-inline", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -5061,6 +5595,8 @@ { "name": "persistent-layer-fused-decode", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -5081,6 +5617,8 @@ { "name": "persistent-rmsnorm-qkv-fused", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -5101,6 +5639,8 @@ { "name": "persistent-scheduler-loadbalance", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -5121,6 +5661,8 @@ { "name": "pipeline-microbatch-stash", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 410.2, "gpus": 1, @@ -5142,6 +5684,8 @@ { "name": "powersgd-lowrank-update", "family": "Training, optimizer & RL", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "GB/s", "roofline_us": 392.3, "gpus": 1, @@ -5162,6 +5706,8 @@ { "name": "prefix-cache-block-hash", "family": "KV cache & paging", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 339.7, "gpus": 1, @@ -5181,6 +5727,8 @@ { "name": "prefix-cache-copy-on-write", "family": "KV cache & paging", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 478.7, "gpus": 1, @@ -5201,6 +5749,8 @@ { "name": "prefix-lm-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 834.5, "gpus": 1, @@ -5219,6 +5769,8 @@ { "name": "pyramid-kv-video-attn", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", "metric": "TFLOP/s", "roofline_us": 610.3, "gpus": 1, @@ -5239,6 +5791,8 @@ { "name": "qk-clip-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -5256,6 +5810,8 @@ { "name": "qk-norm-rope-kvwrite", "family": "KV cache & paging", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 534.1, "gpus": 1, @@ -5275,6 +5831,8 @@ { "name": "qlora-nf4-dequant-gemm", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -5294,6 +5852,8 @@ { "name": "quantized-optimizer-state", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 931.3, "gpus": 1, @@ -5314,6 +5874,8 @@ { "name": "qwen-vl-window-attention", "family": "Multimodal & audio", + "tier": "T3", + "tier_why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", "metric": "TFLOP/s", "roofline_us": 570.9, "gpus": 1, @@ -5333,6 +5895,8 @@ { "name": "qwen3-moe-route-align", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 330.9, "gpus": 1, @@ -5353,6 +5917,8 @@ { "name": "qwen3-next-gated-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 2061.6, "gpus": 1, @@ -5372,6 +5938,8 @@ { "name": "qwen3-next-gated-attn-decode", "family": "KV cache & paging", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "GB/s", "roofline_us": 633.5, "gpus": 1, @@ -5392,6 +5960,8 @@ { "name": "qwen3-next-gdn-inproj", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 2368.4, "gpus": 1, @@ -5412,6 +5982,8 @@ { "name": "qwen3-next-gdn-inproj-backward", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 3552.6, "gpus": 1, @@ -5432,6 +6004,8 @@ { "name": "qwen3-next-output-gate-backward", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 4417.7, "gpus": 1, @@ -5453,6 +6027,8 @@ { "name": "radial-attention-backward", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 736.3, "gpus": 1, @@ -5473,6 +6049,8 @@ { "name": "radial-attention-decay", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 4804.2, "gpus": 1, @@ -5492,6 +6070,8 @@ { "name": "radix-prefix-match", "family": "KV cache & paging", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 442.9, "gpus": 1, @@ -5510,6 +6090,8 @@ { "name": "repetition-penalty-logits", "family": "Sampling & speculative decoding", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 625.0, "gpus": 1, @@ -5529,6 +6111,8 @@ { "name": "residual-diff-l1-reduce", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 300.4, "gpus": 1, @@ -5546,6 +6130,8 @@ { "name": "retention-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 331.3, "gpus": 1, @@ -5563,6 +6149,8 @@ { "name": "retention-forward", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "TFLOP/s", "roofline_us": 441.8, "gpus": 1, @@ -5579,6 +6167,8 @@ { "name": "rmsnorm-backward", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 468.8, "gpus": 1, @@ -5596,6 +6186,8 @@ { "name": "rmsnorm-quant-transpose-fused", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 476.1, "gpus": 1, @@ -5615,6 +6207,8 @@ { "name": "rolling-window-video-kv", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 2356.1, "gpus": 1, @@ -5634,6 +6228,8 @@ { "name": "rope-backward", "family": "Training, optimizer & RL", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 722.7, "gpus": 1, @@ -5651,6 +6247,8 @@ { "name": "rope-fused-kv-write", "family": "KV cache & paging", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 312.7, "gpus": 1, @@ -5668,6 +6266,8 @@ { "name": "rwkv7-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 515.4, "gpus": 1, @@ -5685,6 +6285,8 @@ { "name": "rwkv7-forward", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "TFLOP/s", "roofline_us": 417.2, "gpus": 1, @@ -5702,6 +6304,8 @@ { "name": "rwkv7-token-shift-mix", "family": "Linear attention & SSM", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 820.3, "gpus": 1, @@ -5720,6 +6324,8 @@ { "name": "sandwich-norm-residual-fused", "family": "Normalization, RoPE & elementwise fusion", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 937.5, "gpus": 1, @@ -5739,6 +6345,8 @@ { "name": "scheduler-multistep-buffer", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 395.6, "gpus": 1, @@ -5757,6 +6365,8 @@ { "name": "sd3-joint-attention-tripletext", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 4846.0, "gpus": 1, @@ -5777,6 +6387,8 @@ { "name": "sd3-mmdit-block", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 7395.2, "gpus": 1, @@ -5800,6 +6412,8 @@ { "name": "sd3-pos-embed-interpolate", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "GB/s", "roofline_us": 388.6, "gpus": 1, @@ -5820,6 +6434,8 @@ { "name": "sd3-qkv-proj-qknorm", "family": "Image generation \u2014 FLUX / SD3 MMDiT", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 2675.5, "gpus": 1, @@ -5840,6 +6456,8 @@ { "name": "seqparallel-partial-norm", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 390.6, "gpus": 1, @@ -5861,6 +6479,8 @@ { "name": "sequence-packing-loss-mask", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 1223.1, "gpus": 1, @@ -5880,6 +6500,8 @@ { "name": "sequence-unpad-repad", "family": "KV cache & paging", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 863.6, "gpus": 1, @@ -5899,6 +6521,8 @@ { "name": "siglip-attention-pooled", "family": "Multimodal & audio", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 386.9, "gpus": 1, @@ -5918,6 +6542,8 @@ { "name": "simple-gla-forward", "family": "Linear attention & SSM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "TFLOP/s", "roofline_us": 589.0, "gpus": 1, @@ -5935,6 +6561,8 @@ { "name": "sliding-tile-attention-3d", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 4969.9, "gpus": 1, @@ -5955,6 +6583,8 @@ { "name": "sliding-window-kv-evict", "family": "KV cache & paging", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 514.0, "gpus": 1, @@ -5973,6 +6603,8 @@ { "name": "sliding-window-sink-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 368.2, "gpus": 1, @@ -5990,6 +6622,8 @@ { "name": "smoothquant-fused-scale", "family": "Quantization & low-precision GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 312.7, "gpus": 1, @@ -6008,6 +6642,8 @@ { "name": "snake-antialias-activation", "family": "Multimodal & audio", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "GB/s", "roofline_us": 416.7, "gpus": 1, @@ -6028,6 +6664,8 @@ { "name": "softcap-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 785.4, "gpus": 1, @@ -6045,6 +6683,8 @@ { "name": "sparse-attn-backward-3d", "family": "Training, optimizer & RL", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 1058.4, "gpus": 1, @@ -6065,6 +6705,8 @@ { "name": "sparse-attn-mask-build-3d", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "GB/s", "roofline_us": 567.6, "gpus": 1, @@ -6085,6 +6727,8 @@ { "name": "sparse-block-worklist-build", "family": "Video \u2014 sparse / efficient attention", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 282.8, "gpus": 1, @@ -6105,6 +6749,8 @@ { "name": "sparse24-backward-densify", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 463.9, "gpus": 1, @@ -6126,6 +6772,8 @@ { "name": "sparse24-compress-layout", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 366.2, "gpus": 1, @@ -6147,6 +6795,8 @@ { "name": "sparse24-prune-metadata", "family": "Training, optimizer & RL", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 366.2, "gpus": 1, @@ -6167,6 +6817,8 @@ { "name": "sparse24-sparse-gemm", "family": "Training, optimizer & RL", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 883.5, "gpus": 1, @@ -6187,6 +6839,8 @@ { "name": "spatial-upsample-pixelshuffle3d", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T2", + "tier_why": "pixel-shuffle upsample is a layout transform, not a convolution cuDNN accelerates", "metric": "TFLOP/s", "roofline_us": 398.6, "gpus": 1, @@ -6208,6 +6862,8 @@ { "name": "spatial-window-temporal-global", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", "metric": "TFLOP/s", "roofline_us": 1751.5, "gpus": 1, @@ -6227,6 +6883,8 @@ { "name": "spec-decode-tree-attention", "family": "Sampling & speculative decoding", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 791.5, "gpus": 1, @@ -6243,6 +6901,8 @@ { "name": "spec-decode-verify-accept", "family": "Sampling & speculative decoding", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 468.8, "gpus": 1, @@ -6261,6 +6921,8 @@ { "name": "speculative-draft-kv-rollback", "family": "KV cache & paging", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 466.8, "gpus": 1, @@ -6281,6 +6943,8 @@ { "name": "sta-backward", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 5177.0, "gpus": 1, @@ -6300,6 +6964,8 @@ { "name": "sta-ragged-tile-boundary", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 570.2, "gpus": 1, @@ -6320,6 +6986,8 @@ { "name": "step-distill-consistency", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 494.5, "gpus": 1, @@ -6338,6 +7006,8 @@ { "name": "streaming-llm-attention", "family": "KV cache & paging", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "GB/s", "roofline_us": 429.7, "gpus": 1, @@ -6358,6 +7028,8 @@ { "name": "svg-head-classify-route", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 2635.9, "gpus": 1, @@ -6377,6 +7049,8 @@ { "name": "swiglu-backward", "family": "Training, optimizer & RL", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 520.8, "gpus": 1, @@ -6394,6 +7068,8 @@ { "name": "swiglu-blockwise-fp8-quant", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 589.6, "gpus": 1, @@ -6413,6 +7089,8 @@ { "name": "teacache-residual-decision", "family": "Diffusion \u2014 sampling, scheduling, caching", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 593.4, "gpus": 1, @@ -6431,6 +7109,8 @@ { "name": "temporal-downsample-strided", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 414.2, "gpus": 1, @@ -6453,6 +7133,8 @@ { "name": "temporal-strided-attention", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 2761.1, "gpus": 1, @@ -6472,6 +7154,8 @@ { "name": "temporal-upsample-3d", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 517.7, "gpus": 1, @@ -6493,6 +7177,8 @@ { "name": "titans-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 441.8, "gpus": 1, @@ -6509,6 +7195,8 @@ { "name": "titans-forward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 294.5, "gpus": 1, @@ -6525,6 +7213,8 @@ { "name": "trilinear-interpolate-3d", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 371.3, "gpus": 1, @@ -6545,6 +7235,8 @@ { "name": "ttt-linear-backward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 441.8, "gpus": 1, @@ -6561,6 +7253,8 @@ { "name": "ttt-linear-forward", "family": "Linear attention & SSM", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 392.7, "gpus": 1, @@ -6577,6 +7271,8 @@ { "name": "vae-attention-block-highres", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 958.5, "gpus": 1, @@ -6598,6 +7294,8 @@ { "name": "vae-decode-fp8", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 660.1, "gpus": 1, @@ -6620,6 +7318,8 @@ { "name": "vae-decode-tile-feather", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 564.8, "gpus": 1, @@ -6642,6 +7342,8 @@ { "name": "vae-encoder-downblock-fused", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 706.7, "gpus": 1, @@ -6665,6 +7367,8 @@ { "name": "vae-quant-conv", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 511.3, "gpus": 1, @@ -6686,6 +7390,8 @@ { "name": "vae-resblock-fused", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T3", + "tier_why": "convolution with custom padding/grouping/tiling: cuDNN's path for this case is not the fast one, so the tiled kernel is the task", "metric": "TFLOP/s", "roofline_us": 1092.0, "gpus": 1, @@ -6708,6 +7414,8 @@ { "name": "varlen-packed-attention", "family": "Attention \u2014 text LLM", + "tier": "T3", + "tier_why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", "metric": "TFLOP/s", "roofline_us": 1156.9, "gpus": 1, @@ -6726,6 +7434,8 @@ { "name": "video-attn-active-query-refresh", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 417.2, "gpus": 1, @@ -6746,6 +7456,8 @@ { "name": "video-attn-head-adaptive-window", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 709.4, "gpus": 1, @@ -6766,6 +7478,8 @@ { "name": "video-attn-quantized-sparse", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 883.5, "gpus": 1, @@ -6786,6 +7500,8 @@ { "name": "video-attn-sink-frames", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 3926.8, "gpus": 1, @@ -6805,6 +7521,8 @@ { "name": "video-attn-topk-block-select", "family": "Video \u2014 sparse / efficient attention", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 3926.8, "gpus": 1, @@ -6824,6 +7542,8 @@ { "name": "video-cfg-zero-star-rescale", "family": "Video \u2014 sparse / efficient attention", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 304.0, "gpus": 1, @@ -6844,6 +7564,8 @@ { "name": "video-latent-pack-unpack", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 330.4, "gpus": 1, @@ -6865,6 +7587,8 @@ { "name": "vision-patch-embed-2d", "family": "Multimodal & audio", + "tier": "T4", + "tier_why": "dense GEMM: cuBLAS/CUTLASS are already at the hardware limit", "metric": "TFLOP/s", "roofline_us": 563.7, "gpus": 1, @@ -6883,6 +7607,8 @@ { "name": "w4a16-dequant-gemm", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", "metric": "TFLOP/s", "roofline_us": 392.7, "gpus": 1, @@ -6901,6 +7627,8 @@ { "name": "wan-adaln-modulation-backward", "family": "Video \u2014 Wan DiT", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 901.2, "gpus": 1, @@ -6921,6 +7649,8 @@ { "name": "wan-adaln-modulation-fused", "family": "Video \u2014 Wan DiT", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 600.8, "gpus": 1, @@ -6940,6 +7670,8 @@ { "name": "wan-attention-backward-video", "family": "Video \u2014 Wan DiT", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 8958.0, "gpus": 1, @@ -6960,6 +7692,8 @@ { "name": "wan-cross-attention-backward", "family": "Video \u2014 Wan DiT", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 5662.3, "gpus": 1, @@ -6980,6 +7714,8 @@ { "name": "wan-cross-attention-text", "family": "Video \u2014 Wan DiT", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 3401.8, "gpus": 1, @@ -7000,6 +7736,8 @@ { "name": "wan-ffn-gelu-video", "family": "Video \u2014 Wan DiT", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "TFLOP/s", "roofline_us": 13249.8, "gpus": 1, @@ -7021,6 +7759,8 @@ { "name": "wan-moe-expert-switch", "family": "Video \u2014 Wan DiT", + "tier": "T3", + "tier_why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", "metric": "TFLOP/s", "roofline_us": 11777.6, "gpus": 1, @@ -7041,6 +7781,8 @@ { "name": "wan-patch-embed-3dconv", "family": "Video \u2014 Wan DiT", + "tier": "T4", + "tier_why": "plain convolution: cuDNN applies directly and already runs near roofline", "metric": "GB/s", "roofline_us": 608.5, "gpus": 1, @@ -7060,6 +7802,8 @@ { "name": "wan-qkv-proj-qknorm-rope", "family": "Video \u2014 Wan DiT", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 16986.9, "gpus": 1, @@ -7081,6 +7825,8 @@ { "name": "wan-residual-gate-add", "family": "Video \u2014 Wan DiT", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 1201.6, "gpus": 1, @@ -7101,6 +7847,8 @@ { "name": "wan-rope3d-apply", "family": "Video \u2014 Wan DiT", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 1201.6, "gpus": 1, @@ -7120,6 +7868,8 @@ { "name": "wan-sage-qk-int8-quant", "family": "Video \u2014 Wan DiT", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 901.2, "gpus": 1, @@ -7140,6 +7890,8 @@ { "name": "wan-spatiotemporal-attention", "family": "Video \u2014 Wan DiT", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 7166.4, "gpus": 1, @@ -7160,6 +7912,8 @@ { "name": "wan-sta-tile-permute", "family": "Video \u2014 Wan DiT", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 600.8, "gpus": 1, @@ -7179,6 +7933,8 @@ { "name": "wan-unpatchify-head", "family": "Video \u2014 Wan DiT", + "tier": "T3", + "tier_why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", "metric": "GB/s", "roofline_us": 608.5, "gpus": 1, @@ -7199,6 +7955,8 @@ { "name": "warp-specialized-gemv", "family": "Megakernel \u2014 whole-model fusion & primitives", + "tier": "T3", + "tier_why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", "metric": "tokens/s", "roofline_us": null, "gpus": 1, @@ -7218,6 +7976,8 @@ { "name": "wavelet-transform-3d", "family": "Video \u2014 3D causal VAE / tokenizer", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 373.5, "gpus": 1, @@ -7238,6 +7998,8 @@ { "name": "whisper-conv-frontend", "family": "Multimodal & audio", + "tier": "T4", + "tier_why": "plain convolution: cuDNN applies directly and already runs near roofline", "metric": "TFLOP/s", "roofline_us": 566.2, "gpus": 1, @@ -7256,6 +8018,8 @@ { "name": "whisper-cross-attn-decode", "family": "Multimodal & audio", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "GB/s", "roofline_us": 476.8, "gpus": 1, @@ -7275,6 +8039,8 @@ { "name": "yarn-partial-rope", "family": "Normalization, RoPE & elementwise fusion", + "tier": "T2", + "tier_why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", "metric": "GB/s", "roofline_us": 390.7, "gpus": 1, @@ -7292,6 +8058,8 @@ { "name": "zamba-shared-attn-lora-backward", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 1932.7, "gpus": 1, @@ -7313,6 +8081,8 @@ { "name": "zamba-shared-attn-lora-qkv", "family": "Quantization & low-precision GEMM", + "tier": "T3", + "tier_why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", "metric": "TFLOP/s", "roofline_us": 1879.0, "gpus": 1, @@ -7333,6 +8103,8 @@ { "name": "zero3-param-regroup", "family": "MoE \u2014 routing & grouped GEMM", + "tier": "T1", + "tier_why": "elementwise/bandwidth chain: the win is fusing the passes into one", "metric": "GB/s", "roofline_us": 469.2, "gpus": 1,