| outcomes | correct 257 · error 134 · incorrect 8 · compile_fail 1 |
| avg speedup | 17.47x (over 257 correct) |
| avg memory ratio (kernel / ref) | 0.81x (256 samples) |
| avg fusion ratio (ref launches / kernel launches) | 4.84x (256 samples) |
| avg energy ratio (ref / kernel) | 1.41x (215 samples) |
| avg SOL score (0–1) | 0.741 (255 samples) |
| avg DRAM util % | 64.0% |
| avg occupancy % | 177.2% |
| avg max abs error | 5.54e-03 |
| model | n | correct | compiled | avg speedup |
|---|---|---|---|---|
| Kimi-K2.6 | 200 | 110 | 114 | 7.33x |
| gpt-5.5-priority | 200 | 147 | 151 | 25.11x |