| [ |
| { |
| "name": "activation-recompute-fusion", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.3298, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 2356.0963452342858, |
| "ref_metric": 230.8908 |
| }, |
| { |
| "name": "adafactor-factored-update", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0677, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 520.8333333333334, |
| "ref_metric": 325.0389 |
| }, |
| { |
| "name": "adaptive-sparsity-threshold", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0131, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 359.1478429734707, |
| "ref_metric": 63.0037 |
| }, |
| { |
| "name": "alibi-slopes-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0111, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 7.758 |
| }, |
| { |
| "name": "altup-predict-step", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0167, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 651.0456651449203, |
| "ref_metric": 79.9281 |
| }, |
| { |
| "name": "any-res-image-split", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0269, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 336.456298828125, |
| "ref_metric": 128.8881 |
| }, |
| { |
| "name": "async-weight-prefetch-pipeline", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "attention-lse-merge", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0749, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 1057.9427083333333, |
| "ref_metric": 359.4947 |
| }, |
| { |
| "name": "attention-qk-norm-inline", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0141, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 9.8599 |
| }, |
| { |
| "name": "attention-sink-decode", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0086, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 416.6666666666667, |
| "ref_metric": 41.1527 |
| }, |
| { |
| "name": "attn-density-profile-fused", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.011, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1056.964608, |
| "ref_metric": 7.6661 |
| }, |
| { |
| "name": "audio-codec-rvq-quantize", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0202, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 431.41412571428566, |
| "ref_metric": 14.1646 |
| }, |
| { |
| "name": "audio-mel-spectrogram-fused", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0855, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 315.9046173095703, |
| "ref_metric": 410.4547 |
| }, |
| { |
| "name": "awq-dequant-gemm", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0536, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 37.5263 |
| }, |
| { |
| "name": "based-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0022, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 533.2278593828571, |
| "ref_metric": 1.5611 |
| }, |
| { |
| "name": "based-forward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0082, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 480.11598701714286, |
| "ref_metric": 5.7556 |
| }, |
| { |
| "name": "beam-search-step", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0122, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 416.67938232421875, |
| "ref_metric": 58.5811 |
| }, |
| { |
| "name": "block-sparse-video-attn", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0112, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1346.5872910628573, |
| "ref_metric": 7.8659 |
| }, |
| { |
| "name": "blockwise-fp8-gemm", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": 0.0 |
| }, |
| { |
| "name": "cache-hit-skip-gate", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0576, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 593.3761596679688, |
| "ref_metric": 276.7193 |
| }, |
| { |
| "name": "cascade-prefix-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": true, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": 47.2345 |
| }, |
| { |
| "name": "causal-conv1d-decode-step", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0288, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 625.0, |
| "ref_metric": 138.3901 |
| }, |
| { |
| "name": "causal-conv1d-fused", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0198, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 833.3492279052734, |
| "ref_metric": 94.9032 |
| }, |
| { |
| "name": "causal-conv3d-backward", |
| "tier": "T3", |
| "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", |
| "frac": 0.0205, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 728.0113371428571, |
| "ref_metric": 14.3173 |
| }, |
| { |
| "name": "causal-conv3d-cache-step", |
| "tier": "T3", |
| "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", |
| "frac": 0.043, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 465.9272557714285, |
| "ref_metric": 30.1211 |
| }, |
| { |
| "name": "causal-conv3d-forward", |
| "tier": "T3", |
| "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", |
| "frac": 0.0473, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 533.8749805714285, |
| "ref_metric": 33.0775 |
| }, |
| { |
| "name": "causal-conv3d-fp8", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0514, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 640.6499766857143, |
| "ref_metric": 36.0032 |
| }, |
| { |
| "name": "causal-video-kv-cache-decode", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0162, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 2503.3523668114285, |
| "ref_metric": 11.353 |
| }, |
| { |
| "name": "cfg-distill-single-pass", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0588, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 494.4801330566406, |
| "ref_metric": 282.1732 |
| }, |
| { |
| "name": "chunked-prefill-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0122, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 520.7887520914285, |
| "ref_metric": 8.5226 |
| }, |
| { |
| "name": "chunked-prefill-gather", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.208, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 472.1725995962818, |
| "ref_metric": 998.4066 |
| }, |
| { |
| "name": "chunked-zloss-cross-entropy", |
| "tier": "T3", |
| "why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", |
| "frac": 0.2702, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 113092.62457124572, |
| "ref_metric": 189.1689 |
| }, |
| { |
| "name": "clip-logit-scale-fused", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0525, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 36.728 |
| }, |
| { |
| "name": "cogvideox-expert-adaln-block", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0479, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 419.771671295166, |
| "ref_metric": 230.0551 |
| }, |
| { |
| "name": "cogvideox-joint-3d-attention", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0263, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 5546.911615268571, |
| "ref_metric": 18.4283 |
| }, |
| { |
| "name": "cogvideox-shared-ffn-expert-gate", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0643, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 3883.1585455542854, |
| "ref_metric": 44.9937 |
| }, |
| { |
| "name": "cogvideox-sincos-3d-embed", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0716, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 419.771671295166, |
| "ref_metric": 343.5859 |
| }, |
| { |
| "name": "cogvideox-vae-spatial-norm3d", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0187, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 511.77978515625, |
| "ref_metric": 89.8228 |
| }, |
| { |
| "name": "comba-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0017, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 8, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 441.7680647314286, |
| "ref_metric": 1.1779 |
| }, |
| { |
| "name": "comba-decode-step", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0926, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 473.34909439086914, |
| "ref_metric": 444.3857 |
| }, |
| { |
| "name": "comba-forward", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0001, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 220.8840323657143, |
| "ref_metric": 0.0881 |
| }, |
| { |
| "name": "compressed-kv-video-attn", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0266, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1187.2516739657142, |
| "ref_metric": 18.6145 |
| }, |
| { |
| "name": "conformer-conv-module", |
| "tier": "T3", |
| "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", |
| "frac": 0.0405, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": true, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 871.5351771428572, |
| "ref_metric": 28.3327 |
| }, |
| { |
| "name": "conv3d-layout-transform", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0469, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 358.88671875, |
| "ref_metric": 225.026 |
| }, |
| { |
| "name": "conv3d-tiled-decode-blend", |
| "tier": "T3", |
| "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", |
| "frac": 0.0358, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 943.5026929371428, |
| "ref_metric": 25.0856 |
| }, |
| { |
| "name": "cross-attention-varlen", |
| "tier": "T3", |
| "why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", |
| "frac": 0.0256, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 571.0966081828572, |
| "ref_metric": 17.9415 |
| }, |
| { |
| "name": "cross-layer-fusion-2layer", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 8, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "ddim-inversion-step", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.1082, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 593.3761596679688, |
| "ref_metric": 519.2024 |
| }, |
| { |
| "name": "deepseek-mla-lowrank-proj", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0539, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1237.8709313828572, |
| "ref_metric": 37.7497 |
| }, |
| { |
| "name": "deepseek-mla-proj-backward", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0426, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 2830.0766646857146, |
| "ref_metric": 29.8194 |
| }, |
| { |
| "name": "deepseek-mla-vabsorb-outproj", |
| "tier": "T3", |
| "why": "MLA V-absorb folds the value projection into the output projection: a fused GEMM pair, not a plain one", |
| "frac": 0.0579, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 736.2801078857143, |
| "ref_metric": 40.5305 |
| }, |
| { |
| "name": "deepseek-node-dispatch-pack", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.2865, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 456.0343556416531, |
| "ref_metric": 1375.1752 |
| }, |
| { |
| "name": "deepseek-v32-indexer-decode", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0335, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 314.6203358968099, |
| "ref_metric": 160.7546 |
| }, |
| { |
| "name": "deepseek-v32-sparse-mla-decode", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0257, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 570.6787109375, |
| "ref_metric": 123.543 |
| }, |
| { |
| "name": "delta-product-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0007, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 8, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 398.81839177142854, |
| "ref_metric": 0.5091 |
| }, |
| { |
| "name": "delta-rule-forward", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0003, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 539.9387457828572, |
| "ref_metric": 0.2231 |
| }, |
| { |
| "name": "deltaformer-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0033, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1570.7308968228572, |
| "ref_metric": 2.3119 |
| }, |
| { |
| "name": "deltaformer-forward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0133, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 3141.4617936457144, |
| "ref_metric": 9.3324 |
| }, |
| { |
| "name": "depthwise-separable-conv3d", |
| "tier": "T3", |
| "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", |
| "frac": 0.0209, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 396.33420288, |
| "ref_metric": 14.6111 |
| }, |
| { |
| "name": "diff-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0148, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1570.7308968228572, |
| "ref_metric": 10.3476 |
| }, |
| { |
| "name": "dist-allgather-gemm-overlap", |
| "tier": "T3", |
| "why": "2-GPU compute/communication overlap: the GEMM is a library call but the overlap schedule is the task", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1570.7308968228572, |
| "ref_metric": null |
| }, |
| { |
| "name": "dist-allreduce-rmsnorm", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 14062.5, |
| "ref_metric": null |
| }, |
| { |
| "name": "dist-moe-a2a-dispatch", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": null, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 4687.5, |
| "ref_metric": null |
| }, |
| { |
| "name": "dist-oneshot-allreduce", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 3750.0, |
| "ref_metric": null |
| }, |
| { |
| "name": "dist-reduce-scatter-rmsnorm", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 8437.5, |
| "ref_metric": null |
| }, |
| { |
| "name": "dist-tp-embedding-allreduce", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 14062.5, |
| "ref_metric": null |
| }, |
| { |
| "name": "distill-kl-logits-fused", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0218, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 1223.14453125, |
| "ref_metric": 104.5081 |
| }, |
| { |
| "name": "dora-magnitude-decompose", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.2518, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 813.1676920685715, |
| "ref_metric": 176.2439 |
| }, |
| { |
| "name": "dpmsolverpp-step", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0478, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 494.4801330566406, |
| "ref_metric": 229.6033 |
| }, |
| { |
| "name": "dpo-loss-fused", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0482, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 1448.974609375, |
| "ref_metric": 231.4997 |
| }, |
| { |
| "name": "draft-tree-build", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0126, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 478.604007512331, |
| "ref_metric": 60.5123 |
| }, |
| { |
| "name": "dsa-sparse-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0061, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 294.5120431542857, |
| "ref_metric": 4.2646 |
| }, |
| { |
| "name": "dyt-dynamic-tanh-backward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0323, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 703.1393051147461, |
| "ref_metric": 155.2228 |
| }, |
| { |
| "name": "dyt-dynamic-tanh-norm", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0482, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 468.75953674316406, |
| "ref_metric": 231.458 |
| }, |
| { |
| "name": "embedding-backward-scatter", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0542, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 1041.8701171875, |
| "ref_metric": 260.3287 |
| }, |
| { |
| "name": "entropy-bonus-reduce", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0234, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 1223.14453125, |
| "ref_metric": 112.5576 |
| }, |
| { |
| "name": "feature-cache-blend", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0524, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 741.7201995849609, |
| "ref_metric": 251.3262 |
| }, |
| { |
| "name": "flash-attn-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": 0.0 |
| }, |
| { |
| "name": "flex-blockmask-attention", |
| "tier": "T3", |
| "why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", |
| "frac": 0.0081, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 583.0801583542857, |
| "ref_metric": 5.6462 |
| }, |
| { |
| "name": "flow-match-euler-step", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0709, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 494.4801330566406, |
| "ref_metric": 340.4479 |
| }, |
| { |
| "name": "flow-match-loss-backward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0555, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 544.0735816955566, |
| "ref_metric": 266.289 |
| }, |
| { |
| "name": "flux-attention-backward", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0333, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 7190.235428571429, |
| "ref_metric": 23.3135 |
| }, |
| { |
| "name": "flux-attn-fp8", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0382, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 5752.188342857143, |
| "ref_metric": 26.7575 |
| }, |
| { |
| "name": "flux-double-stream-block", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0519, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 8, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 10478.18628534857, |
| "ref_metric": 36.3505 |
| }, |
| { |
| "name": "flux-dual-stream-mlp", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0629, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 5522.1008091428575, |
| "ref_metric": 44.051 |
| }, |
| { |
| "name": "flux-final-layer-unpatchify", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0195, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 478.515625, |
| "ref_metric": 93.4332 |
| }, |
| { |
| "name": "flux-fused-qkv-mlp-proj", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0607, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 4831.838208, |
| "ref_metric": 42.4919 |
| }, |
| { |
| "name": "flux-joint-attention", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0385, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 5752.188342857143, |
| "ref_metric": 26.9205 |
| }, |
| { |
| "name": "flux-modulation-backward", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0321, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 395.5078125, |
| "ref_metric": 154.0267 |
| }, |
| { |
| "name": "flux-modulation-fused", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0568, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 366.2109375, |
| "ref_metric": 272.4412 |
| }, |
| { |
| "name": "flux-qk-rmsnorm-rope", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0277, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 395.5078125, |
| "ref_metric": 132.7251 |
| }, |
| { |
| "name": "flux-rope2d-apply", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0391, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 366.2109375, |
| "ref_metric": 187.8583 |
| }, |
| { |
| "name": "flux-single-stream-block", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0527, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 10478.18628534857, |
| "ref_metric": 36.8941 |
| }, |
| { |
| "name": "forgetting-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0103, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 7.2105 |
| }, |
| { |
| "name": "forgetting-attn-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0091, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 883.5361294628572, |
| "ref_metric": 6.4016 |
| }, |
| { |
| "name": "fp4-stochastic-round", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0067, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 413.818359375, |
| "ref_metric": 32.213 |
| }, |
| { |
| "name": "fp8-attention-backward", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0159, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1963.4136210285712, |
| "ref_metric": 11.1141 |
| }, |
| { |
| "name": "fp8-blockwise-gemm", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0604, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 42.2638 |
| }, |
| { |
| "name": "fp8-flash-attention", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0143, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 10.0445 |
| }, |
| { |
| "name": "fp8-grouped-gemm-masked", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0726, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 687.19476736, |
| "ref_metric": 50.7868 |
| }, |
| { |
| "name": "fp8-kv-decode-attention", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0052, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 214.84375, |
| "ref_metric": 24.9763 |
| }, |
| { |
| "name": "fp8-per-token-quant-transpose", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0238, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 423.17708333333337, |
| "ref_metric": 114.3638 |
| }, |
| { |
| "name": "fp8-quant-dequant-fused", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0496, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 366.2109375, |
| "ref_metric": 237.8472 |
| }, |
| { |
| "name": "frame-anchor-attention", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0283, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 3779.5712204799997, |
| "ref_metric": 19.7999 |
| }, |
| { |
| "name": "fsdp-allgather-unflatten", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.1541, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 400.10035658876103, |
| "ref_metric": 739.7784 |
| }, |
| { |
| "name": "fsdp-flatten-shard-params", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.402, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 585.1526061693827, |
| "ref_metric": 1929.6087 |
| }, |
| { |
| "name": "fsdp-mixed-precision-cast", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0701, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 453.2120501001676, |
| "ref_metric": 336.6286 |
| }, |
| { |
| "name": "fsdp-reduce-scatter-bucket", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0554, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 568.7585814545553, |
| "ref_metric": 265.7684 |
| }, |
| { |
| "name": "fused-adamw-multitensor", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0655, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 4050.701856613159, |
| "ref_metric": 314.3086 |
| }, |
| { |
| "name": "fused-linear-cross-entropy", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.3054, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 221325.80043044573, |
| "ref_metric": 213.7734 |
| }, |
| { |
| "name": "fused-muon-newton-schulz-batched", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0508, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 6626.520970971428, |
| "ref_metric": 35.531 |
| }, |
| { |
| "name": "fused-qk-norm-rope", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0281, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 361.32822434107464, |
| "ref_metric": 134.6813 |
| }, |
| { |
| "name": "fused-residual-rmsnorm", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0678, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 312.50476837158203, |
| "ref_metric": 325.4356 |
| }, |
| { |
| "name": "fused-rmsnorm-fp8-quant", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0245, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 473.6328125, |
| "ref_metric": 117.4205 |
| }, |
| { |
| "name": "fused-topk-topp-filter", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0121, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 416.6698455810547, |
| "ref_metric": 58.1855 |
| }, |
| { |
| "name": "gae-advantage-scan", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0009, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 546.875, |
| "ref_metric": 4.3194 |
| }, |
| { |
| "name": "gated-delta-product-forward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0003, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 478.58207012571427, |
| "ref_metric": 0.2277 |
| }, |
| { |
| "name": "gated-deltanet-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 8, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": 1.5324 |
| }, |
| { |
| "name": "gated-deltanet-decode-step", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0926, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 472.43356704711914, |
| "ref_metric": 444.6262 |
| }, |
| { |
| "name": "gated-deltanet-forward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 8, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": 0.0 |
| }, |
| { |
| "name": "geglu-backward-fused", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.022, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 781.2754313151042, |
| "ref_metric": 105.7764 |
| }, |
| { |
| "name": "gla-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0008, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 441.7680647314286, |
| "ref_metric": 0.5921 |
| }, |
| { |
| "name": "gla-decode-step", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.1227, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 474.2431640625, |
| "ref_metric": 589.0488 |
| }, |
| { |
| "name": "gla-forward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0005, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 441.7680647314286, |
| "ref_metric": 0.334 |
| }, |
| { |
| "name": "gptq-actorder-gemm", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.057, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 39.9084 |
| }, |
| { |
| "name": "gqa-decode-attention", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.036, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 833.3333333333334, |
| "ref_metric": 172.887 |
| }, |
| { |
| "name": "gqa-paged-decode-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0232, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 626.373291015625, |
| "ref_metric": 111.225 |
| }, |
| { |
| "name": "grad-global-norm-clip", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0794, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 2842.398484547933, |
| "ref_metric": 381.1733 |
| }, |
| { |
| "name": "grad-topk-error-feedback", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0462, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 319.0104166666667, |
| "ref_metric": 221.5401 |
| }, |
| { |
| "name": "gradient-accumulation-fused", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.1379, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 2095.4757928848267, |
| "ref_metric": 661.882 |
| }, |
| { |
| "name": "grammar-jump-ahead-compile", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0113, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 417.3787434895833, |
| "ref_metric": 54.1914 |
| }, |
| { |
| "name": "gridwide-barrier-primitive", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "grouped-fp8-gemm", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": 0.0 |
| }, |
| { |
| "name": "groupnorm3d-backward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0192, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 418.701171875, |
| "ref_metric": 92.1255 |
| }, |
| { |
| "name": "groupnorm3d-forward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0309, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 370.27994791666663, |
| "ref_metric": 148.2469 |
| }, |
| { |
| "name": "grpo-advantage-whiten", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0625, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 325.5208333333333, |
| "ref_metric": 299.8599 |
| }, |
| { |
| "name": "grpo-logprob-ratio", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0378, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 1223.14453125, |
| "ref_metric": 181.6374 |
| }, |
| { |
| "name": "gsa-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0009, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 8, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 441.7680647314286, |
| "ref_metric": 0.6484 |
| }, |
| { |
| "name": "gsa-forward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0002, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 392.6827242057143, |
| "ref_metric": 0.1697 |
| }, |
| { |
| "name": "guided-decoding-mask", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0849, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 317.3828125, |
| "ref_metric": 407.682 |
| }, |
| { |
| "name": "hgrn2-forward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0005, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 441.7680647314286, |
| "ref_metric": 0.3181 |
| }, |
| { |
| "name": "hierarchical-coarse-fine-attn", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0166, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 2278.967588571428, |
| "ref_metric": 11.6272 |
| }, |
| { |
| "name": "hifigan-mrf-resblock", |
| "tier": "T3", |
| "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", |
| "frac": 0.1416, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 517.6969508571428, |
| "ref_metric": 99.1042 |
| }, |
| { |
| "name": "hunyuan-attention-backward", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0266, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 9498.157196434286, |
| "ref_metric": 18.6427 |
| }, |
| { |
| "name": "hunyuan-attn-fp8", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0382, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 10602.433553554285, |
| "ref_metric": 26.7682 |
| }, |
| { |
| "name": "hunyuan-dualstream-attn-proj", |
| "tier": "T3", |
| "why": "two per-stream output projections plus per-sample gates: a 2-group GEMM with M=1e5 against M=1e2", |
| "frac": 0.0601, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 2222.6455756799996, |
| "ref_metric": 42.0512 |
| }, |
| { |
| "name": "hunyuan-dualstream-block", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0515, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 8, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 10013.409467245714, |
| "ref_metric": 36.0544 |
| }, |
| { |
| "name": "hunyuan-final-layer-unpatchify", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0235, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 433.7131977081299, |
| "ref_metric": 112.7753 |
| }, |
| { |
| "name": "hunyuan-joint-attention", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0393, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 10657.726464, |
| "ref_metric": 27.522 |
| }, |
| { |
| "name": "hunyuan-mmdit-modulation", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0509, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 284.46197509765625, |
| "ref_metric": 244.1265 |
| }, |
| { |
| "name": "hunyuan-modulation-backward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0309, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 585.9375, |
| "ref_metric": 148.3044 |
| }, |
| { |
| "name": "hunyuan-qk-norm-fused", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0547, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 853.3859252929688, |
| "ref_metric": 262.4447 |
| }, |
| { |
| "name": "hunyuan-rope3d-apply", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0316, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 643.0912017822266, |
| "ref_metric": 151.4511 |
| }, |
| { |
| "name": "hunyuan-singlestream-block", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0527, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 10013.409467245714, |
| "ref_metric": 36.8809 |
| }, |
| { |
| "name": "hunyuan-split-ffn", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0652, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 9939.781456457144, |
| "ref_metric": 45.625 |
| }, |
| { |
| "name": "hunyuan-split-ffn-backward", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0607, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 7951.825165165714, |
| "ref_metric": 42.5032 |
| }, |
| { |
| "name": "hunyuan-token-replace-i2v", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0512, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 566.4825439453125, |
| "ref_metric": 245.6268 |
| }, |
| { |
| "name": "hunyuan-varlen-multiresolution", |
| "tier": "T3", |
| "why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", |
| "frac": 0.0255, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 9484.700607634286, |
| "ref_metric": 17.8453 |
| }, |
| { |
| "name": "hybrid-layer-output-gate", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0619, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1767.0722589257143, |
| "ref_metric": 43.3355 |
| }, |
| { |
| "name": "hybrid-ssm-attn-parallel-decode", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0247, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 731.3569386800131, |
| "ref_metric": 118.5879 |
| }, |
| { |
| "name": "image-token-merge-pixelunshuffle", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0725, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 421.142578125, |
| "ref_metric": 348.0133 |
| }, |
| { |
| "name": "instruction-interpreter-dispatch", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "int4-kv-cache-quant", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0173, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 546.875, |
| "ref_metric": 83.2316 |
| }, |
| { |
| "name": "int4-weight-only-decode-gemv", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.004, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 218.20704142252606, |
| "ref_metric": 19.1005 |
| }, |
| { |
| "name": "int8-w8a8-gemm", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.073, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 51.0749 |
| }, |
| { |
| "name": "istft-overlap-add", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0212, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 516.661008199056, |
| "ref_metric": 101.8793 |
| }, |
| { |
| "name": "kda-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0017, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 8, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 539.9387457828572, |
| "ref_metric": 1.1689 |
| }, |
| { |
| "name": "kda-decode-step", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0926, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 474.25031661987305, |
| "ref_metric": 444.5032 |
| }, |
| { |
| "name": "kda-forward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0046, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 8, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 539.9387457828572, |
| "ref_metric": 3.2374 |
| }, |
| { |
| "name": "kv-block-pack-for-transfer", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0454, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 586.536726138244, |
| "ref_metric": 218.0431 |
| }, |
| { |
| "name": "kv-cache-4bit-groupwise", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0026, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 520.8333333333334, |
| "ref_metric": 12.6002 |
| }, |
| { |
| "name": "kv-cache-block-compaction", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.1288, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 625.0381469726562, |
| "ref_metric": 618.4209 |
| }, |
| { |
| "name": "kv-cache-fp8-quant-paged", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0314, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 473.6328125, |
| "ref_metric": 150.486 |
| }, |
| { |
| "name": "kv-cache-page-alloc", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0564, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 417.9287243944903, |
| "ref_metric": 270.757 |
| }, |
| { |
| "name": "kv-layout-contig-to-paged", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.1294, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 636.1883870946865, |
| "ref_metric": 621.0023 |
| }, |
| { |
| "name": "kv-repage-block-size-convert", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0694, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 645.2573339144388, |
| "ref_metric": 333.0287 |
| }, |
| { |
| "name": "latent-blend-i2v-mask", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0521, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 509.93263721466064, |
| "ref_metric": 250.1366 |
| }, |
| { |
| "name": "latent-interpolate-keyframe", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.019, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 314.7125244140625, |
| "ref_metric": 90.995 |
| }, |
| { |
| "name": "latent-normalize-perchannel", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.072, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 385.52284240722656, |
| "ref_metric": 345.367 |
| }, |
| { |
| "name": "latent-patchify-3d", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.1973, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 300.02593994140625, |
| "ref_metric": 946.9542 |
| }, |
| { |
| "name": "laurel-lowrank-residual", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0308, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 520.9875106811523, |
| "ref_metric": 147.9671 |
| }, |
| { |
| "name": "layernorm-residual-fp8-backward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0289, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 823.974609375, |
| "ref_metric": 138.6289 |
| }, |
| { |
| "name": "layerscale-residual-fused", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.2129, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 703.1345367431641, |
| "ref_metric": 1021.9772 |
| }, |
| { |
| "name": "lightning-attn-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0007, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 441.7680647314286, |
| "ref_metric": 0.4892 |
| }, |
| { |
| "name": "lightning-attn-forward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0005, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 343.59738368, |
| "ref_metric": 0.325 |
| }, |
| { |
| "name": "log-linear-attn-backward", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0056, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 441.7680647314286, |
| "ref_metric": 3.9116 |
| }, |
| { |
| "name": "log-linear-attn-forward", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0132, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 441.7680647314286, |
| "ref_metric": 9.2686 |
| }, |
| { |
| "name": "logits-gather-lastpos", |
| "tier": "T3", |
| "why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", |
| "frac": 0.0535, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1570.7308968228572, |
| "ref_metric": 37.4238 |
| }, |
| { |
| "name": "lora-backward-fused", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.324, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 822.1794538057143, |
| "ref_metric": 226.7667 |
| }, |
| { |
| "name": "lora-fused-gemm", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.3716, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 800.7046173257143, |
| "ref_metric": 260.09 |
| }, |
| { |
| "name": "ltx-fractional-rope3d", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.019, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 510.25390624999994, |
| "ref_metric": 91.3329 |
| }, |
| { |
| "name": "ltx-highcompression-vae-decode", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.189, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 942.2084505600001, |
| "ref_metric": 132.3293 |
| }, |
| { |
| "name": "ltx-per-token-adanorm", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0569, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 429.1534423828125, |
| "ref_metric": 273.2629 |
| }, |
| { |
| "name": "ltx-per-token-adanorm-backward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0492, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 333.7860107421875, |
| "ref_metric": 236.1666 |
| }, |
| { |
| "name": "ltx-rectified-flow-shift", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0505, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 300.94385147094727, |
| "ref_metric": 242.5323 |
| }, |
| { |
| "name": "mamba2-ssd-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0008, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 343.59738368, |
| "ref_metric": 0.5566 |
| }, |
| { |
| "name": "mamba2-ssd-decode-step", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.1222, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 474.2717742919922, |
| "ref_metric": 586.4182 |
| }, |
| { |
| "name": "mamba2-ssd-forward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": 0.0006 |
| }, |
| { |
| "name": "megakernel-batch4-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 9, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-gated-deltanet-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 19, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-gqa-paged-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 9, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-int4-weight-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 9, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-llama1b-bf16-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 9, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-llama1b-fp8-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 9, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-llama1b-longctx-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 9, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-llama8b-bf16-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 9, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-lm-head-sample-inline", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-mamba-hybrid-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 11, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-mtp-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 11, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-nvfp4-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 9, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-prefill-chunk", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 9, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-qwen3-8b-fp8-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 9, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-qwen3-moe-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 10, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-spec-decode-verify", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 9, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "megakernel-untied-head-decode", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 9, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "mesa-net-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0004, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 291.4442093714286, |
| "ref_metric": 0.2888 |
| }, |
| { |
| "name": "mesa-net-forward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0001, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 233.15536749714286, |
| "ref_metric": 0.0694 |
| }, |
| { |
| "name": "min-p-sampling-filter", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0879, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 416.6682561238607, |
| "ref_metric": 421.9222 |
| }, |
| { |
| "name": "mixed-precision-gemm-split", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0718, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 809.9081186742857, |
| "ref_metric": 50.2402 |
| }, |
| { |
| "name": "mla-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0244, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1178.0481726171429, |
| "ref_metric": 17.082 |
| }, |
| { |
| "name": "mla-decode", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": 0.0021 |
| }, |
| { |
| "name": "mla-decode-paged", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0304, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 350.11768341064453, |
| "ref_metric": 145.7798 |
| }, |
| { |
| "name": "mla-prefill", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0289, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1030.79215104, |
| "ref_metric": 20.2525 |
| }, |
| { |
| "name": "mm-embed-merge-scatter", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.2023, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 410.15625, |
| "ref_metric": 970.883 |
| }, |
| { |
| "name": "moba-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0111, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 540.4180948114285, |
| "ref_metric": 7.797 |
| }, |
| { |
| "name": "moba-block-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": 6.2968 |
| }, |
| { |
| "name": "mochi-asymm-joint-attention", |
| "tier": "T3", |
| "why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", |
| "frac": 0.0323, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 5107.665188571428, |
| "ref_metric": 22.6148 |
| }, |
| { |
| "name": "mochi-asymm-qkv-proj", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.061, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 3621.9372924342856, |
| "ref_metric": 42.6655 |
| }, |
| { |
| "name": "mochi-tanh-gated-rmsnorm", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0503, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 429.1534423828125, |
| "ref_metric": 241.4728 |
| }, |
| { |
| "name": "mochi-visual-text-split-mlp", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.065, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 4829.249723245714, |
| "ref_metric": 45.4709 |
| }, |
| { |
| "name": "moe-capacity-drop-dispatch", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0126, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 483.34757486979163, |
| "ref_metric": 60.7125 |
| }, |
| { |
| "name": "moe-combine-backward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0273, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 581.2835693359375, |
| "ref_metric": 131.2667 |
| }, |
| { |
| "name": "moe-combine-weighted", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0267, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 205.17985026041669, |
| "ref_metric": 128.2668 |
| }, |
| { |
| "name": "moe-expert-bias-update", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0048, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 260.467529296875, |
| "ref_metric": 23.0509 |
| }, |
| { |
| "name": "moe-expert-weight-gather", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0873, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 729.1671137015024, |
| "ref_metric": 419.1689 |
| }, |
| { |
| "name": "moe-group-limited-routing", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0048, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 221.35436534881592, |
| "ref_metric": 22.9193 |
| }, |
| { |
| "name": "moe-grouped-gemm-contiguous", |
| "tier": "T3", |
| "why": "grouped GEMM over a contiguous expert layout: no single library call covers it", |
| "frac": 0.0564, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1374.38953472, |
| "ref_metric": 39.4584 |
| }, |
| { |
| "name": "moe-grouped-gemm-varm", |
| "tier": "T3", |
| "why": "variable-M grouped GEMM: cuBLAS has no such call", |
| "frac": 0.0373, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1570.7308968228572, |
| "ref_metric": 26.1395 |
| }, |
| { |
| "name": "moe-grouped-swiglu", |
| "tier": "T3", |
| "why": "grouped GEMM with a fused SwiGLU epilogue", |
| "frac": 0.0539, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1570.7308968228572, |
| "ref_metric": 37.7116 |
| }, |
| { |
| "name": "moe-router-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0425, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 429.49672960000004, |
| "ref_metric": 29.7736 |
| }, |
| { |
| "name": "moe-shared-expert-fused", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0481, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1032.13432832, |
| "ref_metric": 33.6427 |
| }, |
| { |
| "name": "moe-token-permute", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.1298, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 205.18004971866807, |
| "ref_metric": 622.945 |
| }, |
| { |
| "name": "moe-topk-softmax-fused", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0155, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 276.6927083333333, |
| "ref_metric": 74.1755 |
| }, |
| { |
| "name": "mrope-multimodal-apply", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0346, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 325.5208333333333, |
| "ref_metric": 165.9735 |
| }, |
| { |
| "name": "mtp-head-forward", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0607, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 3337.8031557485715, |
| "ref_metric": 42.5029 |
| }, |
| { |
| "name": "mtp-multi-head-loss", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0366, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 1666.6666666666667, |
| "ref_metric": 175.4855 |
| }, |
| { |
| "name": "multi-lora-batched-decode", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0283, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 477.7331999503076, |
| "ref_metric": 135.7824 |
| }, |
| { |
| "name": "muon-newton-schulz", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": 0.0 |
| }, |
| { |
| "name": "mxfp4-dequant-gemm", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0472, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 33.0729 |
| }, |
| { |
| "name": "mxfp4-training-backward", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0607, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 42.4581 |
| }, |
| { |
| "name": "mxfp8-blockwise-gemm", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0595, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 41.668 |
| }, |
| { |
| "name": "mxfp8-dual-axis-quantize", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0253, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 423.17708333333337, |
| "ref_metric": 121.4634 |
| }, |
| { |
| "name": "ngpt-l2norm-residual", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.039, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 703.129768371582, |
| "ref_metric": 187.4287 |
| }, |
| { |
| "name": "ngram-lookahead-match", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0163, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 300.535907347997, |
| "ref_metric": 78.4312 |
| }, |
| { |
| "name": "noise-add-scaled", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.068, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 494.4801330566406, |
| "ref_metric": 326.581 |
| }, |
| { |
| "name": "nsa-compression-branch", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0109, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 393.0654544457143, |
| "ref_metric": 7.6478 |
| }, |
| { |
| "name": "nsa-selected-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": 0.0001 |
| }, |
| { |
| "name": "nsa-selected-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0073, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 392.6827242057143, |
| "ref_metric": 5.094 |
| }, |
| { |
| "name": "nvfp4-dequant-gemm", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0465, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 32.5616 |
| }, |
| { |
| "name": "nvfp4-gemm-backward", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.1296, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 90.6909 |
| }, |
| { |
| "name": "nvfp4-quantize", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0124, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 534.6679695261022, |
| "ref_metric": 59.6044 |
| }, |
| { |
| "name": "optimizer-state-reshard", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.1181, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 802.3745752871037, |
| "ref_metric": 566.7412 |
| }, |
| { |
| "name": "paged-kv-append", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.1385, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 629.8926472663879, |
| "ref_metric": 664.5607 |
| }, |
| { |
| "name": "path-attn-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0008, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 14, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 865.1291267657143, |
| "ref_metric": 0.5947 |
| }, |
| { |
| "name": "path-attn-forward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0007, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 14, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 805.306368, |
| "ref_metric": 0.4908 |
| }, |
| { |
| "name": "penalty-count-bincount-build", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.1308, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 417.88005813335377, |
| "ref_metric": 627.7482 |
| }, |
| { |
| "name": "persistent-kv-append-inline", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "persistent-layer-fused-decode", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 8, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "persistent-rmsnorm-qkv-fused", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "persistent-scheduler-loadbalance", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": null, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": true, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "pipeline-microbatch-stash", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.054, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 410.15625, |
| "ref_metric": 259.1393 |
| }, |
| { |
| "name": "powersgd-lowrank-update", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0948, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 392.303466796875, |
| "ref_metric": 454.95 |
| }, |
| { |
| "name": "prefix-cache-block-hash", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0004, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 339.71020486205816, |
| "ref_metric": 2.1117 |
| }, |
| { |
| "name": "prefix-cache-copy-on-write", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0584, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 478.7498712539673, |
| "ref_metric": 280.5183 |
| }, |
| { |
| "name": "prefix-lm-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0155, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 834.4867401142857, |
| "ref_metric": 10.8847 |
| }, |
| { |
| "name": "pyramid-kv-video-attn", |
| "tier": "T3", |
| "why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", |
| "frac": 0.0185, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 610.3071831771429, |
| "ref_metric": 12.9318 |
| }, |
| { |
| "name": "qk-clip-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0124, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 8.654 |
| }, |
| { |
| "name": "qk-norm-rope-kvwrite", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0313, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 534.0576171875, |
| "ref_metric": 150.1257 |
| }, |
| { |
| "name": "qlora-nf4-dequant-gemm", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.1368, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 95.7762 |
| }, |
| { |
| "name": "quantized-optimizer-state", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0265, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 931.3201904296875, |
| "ref_metric": 127.2957 |
| }, |
| { |
| "name": "qwen-vl-window-attention", |
| "tier": "T3", |
| "why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", |
| "frac": 0.003, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 570.8940361142857, |
| "ref_metric": 2.0799 |
| }, |
| { |
| "name": "qwen3-moe-route-align", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0164, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 330.90909089272225, |
| "ref_metric": 78.7732 |
| }, |
| { |
| "name": "qwen3-next-gated-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0187, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 2061.58430208, |
| "ref_metric": 13.0986 |
| }, |
| { |
| "name": "qwen3-next-gated-attn-decode", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0211, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 633.4940592447916, |
| "ref_metric": 101.466 |
| }, |
| { |
| "name": "qwen3-next-gdn-inproj", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0616, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 2368.3676803657145, |
| "ref_metric": 43.1066 |
| }, |
| { |
| "name": "qwen3-next-gdn-inproj-backward", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0582, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 3552.5515205485713, |
| "ref_metric": 40.7519 |
| }, |
| { |
| "name": "qwen3-next-output-gate-backward", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0518, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 4417.680647314286, |
| "ref_metric": 36.2459 |
| }, |
| { |
| "name": "radial-attention-backward", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0123, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 736.2801078857143, |
| "ref_metric": 8.6305 |
| }, |
| { |
| "name": "radial-attention-decay", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0171, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 4804.227703954286, |
| "ref_metric": 11.9661 |
| }, |
| { |
| "name": "radix-prefix-match", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0014, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 442.9414262995124, |
| "ref_metric": 6.7531 |
| }, |
| { |
| "name": "repetition-penalty-logits", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.084, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 625.0035762786865, |
| "ref_metric": 403.0972 |
| }, |
| { |
| "name": "residual-diff-l1-reduce", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0403, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 300.40740966796875, |
| "ref_metric": 193.4371 |
| }, |
| { |
| "name": "retention-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.001, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 331.3260485485714, |
| "ref_metric": 0.734 |
| }, |
| { |
| "name": "retention-forward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0005, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 441.7680647314286, |
| "ref_metric": 0.3315 |
| }, |
| { |
| "name": "rmsnorm-backward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0337, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 468.7627156575521, |
| "ref_metric": 161.7218 |
| }, |
| { |
| "name": "rmsnorm-quant-transpose-fused", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0182, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 476.07421875, |
| "ref_metric": 87.2002 |
| }, |
| { |
| "name": "rolling-window-video-kv", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0149, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 2356.0963452342858, |
| "ref_metric": 10.4269 |
| }, |
| { |
| "name": "rope-backward", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0444, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 722.65625, |
| "ref_metric": 213.2365 |
| }, |
| { |
| "name": "rope-fused-kv-write", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0405, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 312.7066294352213, |
| "ref_metric": 194.6383 |
| }, |
| { |
| "name": "rwkv7-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0015, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 13, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 515.39607552, |
| "ref_metric": 1.023 |
| }, |
| { |
| "name": "rwkv7-forward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0004, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 417.22539446857144, |
| "ref_metric": 0.3103 |
| }, |
| { |
| "name": "rwkv7-token-shift-mix", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0354, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 820.3268051147461, |
| "ref_metric": 169.7827 |
| }, |
| { |
| "name": "sandwich-norm-residual-fused", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0449, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 937.504768371582, |
| "ref_metric": 215.5592 |
| }, |
| { |
| "name": "scheduler-multistep-buffer", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0353, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 395.5841064453125, |
| "ref_metric": 169.6694 |
| }, |
| { |
| "name": "sd3-joint-attention-tripletext", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0311, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 4846.0049554285715, |
| "ref_metric": 21.786 |
| }, |
| { |
| "name": "sd3-mmdit-block", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.0408, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 8, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 7395.2129536, |
| "ref_metric": 28.5704 |
| }, |
| { |
| "name": "sd3-pos-embed-interpolate", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.025, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 388.59049479166663, |
| "ref_metric": 120.0727 |
| }, |
| { |
| "name": "sd3-qkv-proj-qknorm", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0558, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 2675.522554148571, |
| "ref_metric": 39.0389 |
| }, |
| { |
| "name": "seqparallel-partial-norm", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0726, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 390.625, |
| "ref_metric": 348.4383 |
| }, |
| { |
| "name": "sequence-packing-loss-mask", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.038, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 1223.14453125, |
| "ref_metric": 182.5226 |
| }, |
| { |
| "name": "sequence-unpad-repad", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.2958, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 863.6427678478261, |
| "ref_metric": 1419.9999 |
| }, |
| { |
| "name": "siglip-attention-pooled", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0249, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 386.924544, |
| "ref_metric": 17.4111 |
| }, |
| { |
| "name": "simple-gla-forward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0005, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 589.0240863085714, |
| "ref_metric": 0.3156 |
| }, |
| { |
| "name": "sliding-tile-attention-3d", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0151, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 4969.890728228572, |
| "ref_metric": 10.5973 |
| }, |
| { |
| "name": "sliding-window-kv-evict", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.095, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 514.0247195959091, |
| "ref_metric": 455.9858 |
| }, |
| { |
| "name": "sliding-window-sink-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0014, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 368.15203766857144, |
| "ref_metric": 1.0046 |
| }, |
| { |
| "name": "smoothquant-fused-scale", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0288, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 312.70503997802734, |
| "ref_metric": 138.292 |
| }, |
| { |
| "name": "snake-antialias-activation", |
| "tier": "T3", |
| "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", |
| "frac": 0.0087, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 416.6666666666667, |
| "ref_metric": 41.7805 |
| }, |
| { |
| "name": "softcap-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0111, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 785.3654484114286, |
| "ref_metric": 7.7425 |
| }, |
| { |
| "name": "sparse-attn-backward-3d", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0112, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1058.4026550857143, |
| "ref_metric": 7.8367 |
| }, |
| { |
| "name": "sparse-attn-mask-build-3d", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0472, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 567.626953125, |
| "ref_metric": 226.4278 |
| }, |
| { |
| "name": "sparse-block-worklist-build", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0118, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 282.7978126394252, |
| "ref_metric": 56.6957 |
| }, |
| { |
| "name": "sparse24-backward-densify", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0392, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 463.8671875, |
| "ref_metric": 188.1122 |
| }, |
| { |
| "name": "sparse24-compress-layout", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0147, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 366.2109375, |
| "ref_metric": 70.5673 |
| }, |
| { |
| "name": "sparse24-prune-metadata", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0009, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 366.2109375, |
| "ref_metric": 4.3205 |
| }, |
| { |
| "name": "sparse24-sparse-gemm", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0302, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 883.5361294628572, |
| "ref_metric": 21.1393 |
| }, |
| { |
| "name": "spatial-upsample-pixelshuffle3d", |
| "tier": "T2", |
| "why": "pixel-shuffle upsample is a layout transform, not a convolution cuDNN accelerates", |
| "frac": 0.0347, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 398.62665216, |
| "ref_metric": 24.2893 |
| }, |
| { |
| "name": "spatial-window-temporal-global", |
| "tier": "T3", |
| "why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", |
| "frac": 0.0064, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1751.5413503999998, |
| "ref_metric": 4.4759 |
| }, |
| { |
| "name": "spec-decode-tree-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.01, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 791.5011159771428, |
| "ref_metric": 6.9869 |
| }, |
| { |
| "name": "spec-decode-verify-accept", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0657, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 468.75476837158203, |
| "ref_metric": 315.5627 |
| }, |
| { |
| "name": "speculative-draft-kv-rollback", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0676, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 466.78662300109863, |
| "ref_metric": 324.5761 |
| }, |
| { |
| "name": "sta-backward", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0129, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 5176.969508571429, |
| "ref_metric": 9.0407 |
| }, |
| { |
| "name": "sta-ragged-tile-boundary", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0137, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 570.2443183542856, |
| "ref_metric": 9.5739 |
| }, |
| { |
| "name": "step-distill-consistency", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0652, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 494.4801330566406, |
| "ref_metric": 313.0097 |
| }, |
| { |
| "name": "streaming-llm-attention", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0084, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 429.6875, |
| "ref_metric": 40.3629 |
| }, |
| { |
| "name": "svg-head-classify-route", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0202, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 8, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 2635.9403081142855, |
| "ref_metric": 14.1206 |
| }, |
| { |
| "name": "swiglu-backward", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0503, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 520.8333333333334, |
| "ref_metric": 241.253 |
| }, |
| { |
| "name": "swiglu-blockwise-fp8-quant", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0349, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 589.599609375, |
| "ref_metric": 167.3763 |
| }, |
| { |
| "name": "teacache-residual-decision", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0437, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 593.3761596679688, |
| "ref_metric": 209.7849 |
| }, |
| { |
| "name": "temporal-downsample-strided", |
| "tier": "T3", |
| "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", |
| "frac": 0.0405, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 414.15756068571426, |
| "ref_metric": 28.3249 |
| }, |
| { |
| "name": "temporal-strided-attention", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0071, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 2761.0504045714288, |
| "ref_metric": 4.9653 |
| }, |
| { |
| "name": "temporal-upsample-3d", |
| "tier": "T3", |
| "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", |
| "frac": 0.0401, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 517.6969508571428, |
| "ref_metric": 28.0915 |
| }, |
| { |
| "name": "titans-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0008, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 7, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 441.7680647314286, |
| "ref_metric": 0.5635 |
| }, |
| { |
| "name": "titans-forward", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0003, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 294.5120431542857, |
| "ref_metric": 0.2207 |
| }, |
| { |
| "name": "trilinear-interpolate-3d", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0466, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 371.34806315104163, |
| "ref_metric": 223.6343 |
| }, |
| { |
| "name": "ttt-linear-backward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0008, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 6, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 441.7680647314286, |
| "ref_metric": 0.5847 |
| }, |
| { |
| "name": "ttt-linear-forward", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0043, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 6, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 392.6827242057143, |
| "ref_metric": 3.0383 |
| }, |
| { |
| "name": "vae-attention-block-highres", |
| "tier": "T3", |
| "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", |
| "frac": 0.0465, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": true, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 958.5482605714286, |
| "ref_metric": 32.5739 |
| }, |
| { |
| "name": "vae-decode-fp8", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0482, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 660.0636123428571, |
| "ref_metric": 33.7674 |
| }, |
| { |
| "name": "vae-decode-tile-feather", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0388, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 564.8040771484375, |
| "ref_metric": 186.1219 |
| }, |
| { |
| "name": "vae-encoder-downblock-fused", |
| "tier": "T3", |
| "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", |
| "frac": 0.0374, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 706.6563379199999, |
| "ref_metric": 26.2041 |
| }, |
| { |
| "name": "vae-quant-conv", |
| "tier": "T3", |
| "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", |
| "frac": 0.0307, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 511.3057865142858, |
| "ref_metric": 21.519 |
| }, |
| { |
| "name": "vae-resblock-fused", |
| "tier": "T3", |
| "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", |
| "frac": 0.0388, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1092.0170057142857, |
| "ref_metric": 27.1702 |
| }, |
| { |
| "name": "varlen-packed-attention", |
| "tier": "T3", |
| "why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", |
| "frac": 0.0128, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1156.8772008228573, |
| "ref_metric": 8.9276 |
| }, |
| { |
| "name": "video-attn-active-query-refresh", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0304, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 417.22539446857144, |
| "ref_metric": 21.273 |
| }, |
| { |
| "name": "video-attn-head-adaptive-window", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0025, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 4, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 709.4365622857143, |
| "ref_metric": 1.7646 |
| }, |
| { |
| "name": "video-attn-quantized-sparse", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0202, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 883.5361294628572, |
| "ref_metric": 14.1205 |
| }, |
| { |
| "name": "video-attn-sink-frames", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0247, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 3926.8272420571425, |
| "ref_metric": 17.3186 |
| }, |
| { |
| "name": "video-attn-topk-block-select", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0326, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 3926.8272420571425, |
| "ref_metric": 22.7906 |
| }, |
| { |
| "name": "video-cfg-zero-star-rescale", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0383, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 303.9836883544922, |
| "ref_metric": 183.8145 |
| }, |
| { |
| "name": "video-latent-pack-unpack", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.1679, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 330.4481506347656, |
| "ref_metric": 805.9575 |
| }, |
| { |
| "name": "vision-patch-embed-2d", |
| "tier": "T4", |
| "why": "dense GEMM: cuBLAS/CUTLASS are already at the hardware limit", |
| "frac": 0.0205, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 563.7144576, |
| "ref_metric": 14.3676 |
| }, |
| { |
| "name": "w4a16-dequant-gemm", |
| "tier": "T3", |
| "why": "quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is hand-written -- async copy, double buffering, fragment layouts", |
| "frac": 0.0453, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": true |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 392.6827242057143, |
| "ref_metric": 31.6845 |
| }, |
| { |
| "name": "wan-adaln-modulation-backward", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.027, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 901.2222290039062, |
| "ref_metric": 129.4269 |
| }, |
| { |
| "name": "wan-adaln-modulation-fused", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0384, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 600.8148193359375, |
| "ref_metric": 184.2033 |
| }, |
| { |
| "name": "wan-attention-backward-video", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0111, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 8957.952, |
| "ref_metric": 7.7506 |
| }, |
| { |
| "name": "wan-cross-attention-backward", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0337, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 5, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 5662.3104, |
| "ref_metric": 23.5932 |
| }, |
| { |
| "name": "wan-cross-attention-text", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.036, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 3401.8099199999997, |
| "ref_metric": 25.174 |
| }, |
| { |
| "name": "wan-ffn-gelu-video", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": 0.064, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 13249.806336, |
| "ref_metric": 44.7734 |
| }, |
| { |
| "name": "wan-moe-expert-switch", |
| "tier": "T3", |
| "why": "attention with custom masking/sparsity: no library kernel applies as-is, so the tiled online-softmax pipeline is written by hand", |
| "frac": 0.0624, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": true, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 11777.605631999999, |
| "ref_metric": 43.6683 |
| }, |
| { |
| "name": "wan-patch-embed-3dconv", |
| "tier": "T4", |
| "why": "plain convolution: cuDNN applies directly and already runs near roofline", |
| "frac": 0.0214, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 608.45414797465, |
| "ref_metric": 102.6147 |
| }, |
| { |
| "name": "wan-qkv-proj-qknorm-rope", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.2955, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 16986.9312, |
| "ref_metric": 206.8575 |
| }, |
| { |
| "name": "wan-residual-gate-add", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.051, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 1201.629638671875, |
| "ref_metric": 245.0307 |
| }, |
| { |
| "name": "wan-rope3d-apply", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0367, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 1201.629638671875, |
| "ref_metric": 176.1293 |
| }, |
| { |
| "name": "wan-sage-qk-int8-quant", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0256, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 901.2222290039062, |
| "ref_metric": 122.8106 |
| }, |
| { |
| "name": "wan-spatiotemporal-attention", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0103, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 7166.3616, |
| "ref_metric": 7.1812 |
| }, |
| { |
| "name": "wan-sta-tile-permute", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.2003, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 600.8148193359375, |
| "ref_metric": 961.5004 |
| }, |
| { |
| "name": "wan-unpatchify-head", |
| "tier": "T3", |
| "why": "grouped / variable-M GEMM: there is no single library call for it, so the per-group tiling and scheduling are the task", |
| "frac": 0.0172, |
| "analysis": { |
| "grouped": true, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 1, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 608.4521611531576, |
| "ref_metric": 82.4882 |
| }, |
| { |
| "name": "warp-specialized-gemv", |
| "tier": "T3", |
| "why": "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline", |
| "frac": null, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": true, |
| "epilogue": true, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": null, |
| "roofline_us": null, |
| "ref_metric": null |
| }, |
| { |
| "name": "wavelet-transform-3d", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.0289, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 373.53515625, |
| "ref_metric": 138.5753 |
| }, |
| { |
| "name": "whisper-conv-frontend", |
| "tier": "T4", |
| "why": "plain convolution: cuDNN applies directly and already runs near roofline", |
| "frac": 0.1619, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": true, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": true, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 566.23104, |
| "ref_metric": 113.3154 |
| }, |
| { |
| "name": "whisper-cross-attn-decode", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.036, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 2, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 476.837158203125, |
| "ref_metric": 172.8725 |
| }, |
| { |
| "name": "yarn-partial-rope", |
| "tier": "T2", |
| "why": "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction", |
| "frac": 0.0411, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": true, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "GB/s", |
| "roofline_us": 390.7012939453125, |
| "ref_metric": 197.5155 |
| }, |
| { |
| "name": "zamba-shared-attn-lora-backward", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0304, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1932.7352832, |
| "ref_metric": 21.2649 |
| }, |
| { |
| "name": "zamba-shared-attn-lora-qkv", |
| "tier": "T3", |
| "why": "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat", |
| "frac": 0.0581, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 3, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": false |
| }, |
| "metric": "TFLOP/s", |
| "roofline_us": 1879.048192, |
| "ref_metric": 40.6886 |
| }, |
| { |
| "name": "zero3-param-regroup", |
| "tier": "T1", |
| "why": "elementwise/bandwidth chain: the win is fusing the passes into one", |
| "frac": 0.2053, |
| "analysis": { |
| "grouped": false, |
| "custom_conv": false, |
| "epilogue": false, |
| "n_mm": 0, |
| "sdpa": false, |
| "conv": false, |
| "reduce": false, |
| "masky": false, |
| "quant": true |
| }, |
| "metric": "GB/s", |
| "roofline_us": 469.21253204345703, |
| "ref_metric": 985.2554 |
| } |
| ] |
|
|
