| outcomes | correct 95 · error 57 |
| avg speedup | 1000.22x (over 95 correct) |
| avg memory ratio (kernel / ref) | 1.09x (95 samples) |
| avg fusion ratio (ref launches / kernel launches) | 3620.71x (95 samples) |
| avg energy ratio (ref / kernel) | 15.26x (5 samples) |
| avg SOL score (0–1) | 0.199 (95 samples) |
| avg DRAM util % | 14.2% |
| avg occupancy % | 72.0% |
| avg max abs error | 8.42e-06 |
| model | n | correct | compiled | avg speedup |
|---|---|---|---|---|
| Kimi-K2.6 | 76 | 41 | 41 | 466.40x |
| gpt-5.5-priority | 76 | 54 | 54 | 1405.53x |