Throughput benchmarking on 80 study points synthetic workload configurations (full grid)
28.15Mean RatioTiledAttention
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TiledAttentionComparison Baseline=PyTorch SDPA (math)2026.03 | 28.15 | 20.31 | 100 | |
| TiledAttentionComparison Baseline=standard eager attention2026.03 | 14.36 | 10.76 | 100 | |
| TiledAttentionComparison Baseline=fused PyTorch SDPA2026.03 | 0.632 | 0.634 | 5 |