CUDA Kernel Optimization on ParEval 5-task subset
85Pass@1HPC-Coder-V2 + OptiML-X
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| HPC-Coder-V2 + OptiML-XSelection strategy=best-of-k, Optimization framework=OptiML-X2026.02 | 85 | 100 | 100 | 1.14 | 1.21 | 1.24 | |
| OptiML (Gen + X)Selection strategy=best-of-k2026.02 | 85 | 100 | 100 | 1.18 | 1.26 | 1.3 | |
| GPT 5.1 + OptiML-XSelection strategy=best-of-k, Optimization framework=OptiML-X2026.02 | 80 | 100 | 100 | 1.1 | 1.16 | 1.19 | |
| StarCoder2 + OptiML-XSelection strategy=best-of-k, Optimization framework=OptiML-X2026.02 | 75 | 100 | 100 | 1.06 | 1.12 | 1.17 | |
| Qwen2.5-Coder + OptiML-XSelection strategy=best-of-k, Optimization framework=OptiML-X2026.02 | 70 | 100 | 100 | 1.03 | 1.08 | 1.1 | |
| GPT 5.1Selection strategy=best-of-k2026.02 | 40 | 40 | 40 | 1 | 1 | 1 | |
| Qwen2.5-CoderSelection strategy=best-of-k2026.02 | 20 | 40 | 40 | 0.89 | 0.93 | 0.95 | |
| StarCoder2Selection strategy=best-of-k2026.02 | 20 | 40 | 40 | 0.82 | 0.86 | 0.89 | |
| HPC-Coder-V2Selection strategy=best-of-k2026.02 | 20 | 40 | 40 | 0.86 | 0.9 | 0.92 |