CUDA Kernel Generation on KernelBench Level 1 (single-op)
89fast1 Performance (Native)ALPHALAB
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ALPHALABModel=Opus 4.6, Hardware=H100 NVL2026.03 | 89 | 45 | 82 | |
| ALPHALABModel=GPT-5.2, Hardware=H100 NVL2026.03 | 66 | 61 | 82 | |
| DeepSeek R1Evaluation Protocol=R1 (iter.), Hardware=L40S GPU, Note=10 iterative calls including profiler feedback2026.03 | 43 | — | — | |
| DeepSeek R1Evaluation Protocol=1-shot, Hardware=L40S GPU2026.03 | 12 | — | — | |
| OpenAI o1Evaluation Protocol=1-shot, Hardware=L40S GPU2026.03 | 10 | — | — |