| outcomes | compile_fail 205 · correct 48 · incorrect 14 |
| avg speedup | 218.11x (over 48 correct) |
| avg memory ratio (kernel / ref) | 0.94x (48 samples) |
| avg fusion ratio (ref launches / kernel launches) | 700.93x (48 samples) |
| avg energy ratio (ref / kernel) | 3.95x (11 samples) |
| avg SOL score (0–1) | 0.367 (47 samples) |
| avg DRAM util % | 27.5% |
| avg occupancy % | 235.2% |
| avg max abs error | 2.32e-05 |
| model | n | correct | compiled | avg speedup |
|---|---|---|---|---|
| FW-GLM-5-1 | 89 | 24 | 29 | 254.02x |
| gpt-5.5-priority | 89 | 3 | 4 | 6.27x |
| grok-4.3 | 89 | 21 | 29 | 207.33x |