Join
KernelBench operators
Faster GPU kernels for PyTorch operators, exactly as correct as the reference (placeholder)
| Route | speed-up | Improved | matmul | conv2d | softmax | layernorm | gelu | attention | groupnorm | maxpool | cumsum | argmax | mse-loss | lstm-cell |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Baseline · PyTorch eager | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | |
| autotune-search | 1.74 | 11 / 12 | 1.7 | 3.52 | 3.04 | 1.58 | 2.1 | 1.27 | 0.677 | 1.89 | 2.25 | 1.93 | 1.38 | 1.31 |
| fused-attention | 1.52 | 9 / 12 | 1.81 | 4.32 | 3.44 | 1.43 | 1.74 | 1.43 | 0.859 | 1.22 | 0.659 | 2.13 | 0.784 | 1.38 |
| triton-fusion | 1.38 | 10 / 12 | 2.34 | 3.18 | 2.58 | 1.91 | 1.13 | 1.23 | 0.566 | 1.25 | 0.522 | 1.58 | 1.01 | 1.58 |
| cuda-tiling | 1.29 | 8 / 12 | 1.38 | 1.61 | 0.703 | 2.07 | 2.08 | 0.795 | 0.859 | 1.48 | 2.45 | 1.11 | 1.43 | 0.805 |
| torch-compile | 1.12 | 7 / 12 | 1.24 | 0.939 | 0.862 | 1.08 | 1.49 | 0.869 | 0.989 | 0.96 | 1.44 | 1.34 | 1.09 | 1.4 |