Multi-Head Self-Attention on CUDA-LLM kernels task suite
6.85Execution Time (s)OptiML
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| OptiMLOptimization=G+X, Hardware=A100 80GB2026.02 | 6.85 | 1.15 | 50.28 | 24.88 | 61.79 | 1,064 | 3,961 | 2,012 | |
| HPC-Coder-V2 + OptiML-XOptimization=OptiML-X, Hardware=A100 80GB2026.02 | 7.55 | 1.05 | 49.31 | 25.72 | 60.88 | 1,086 | 4,124 | 2,048 | |
| GPT 5.1 + OptiML-XOptimization=OptiML-X, Hardware=A100 80GB2026.02 | 7.9 | — | 48.62 | 26.41 | 60.32 | 1,117 | 4,213 | 2,094 | |
| StarCoder2 + OptiML-XOptimization=OptiML-X, Hardware=A100 80GB2026.02 | 8.22 | 0.96 | 47.55 | 26.98 | 59.46 | 1,144 | 4,286 | 2,126 | |
| Qwen2.5-Coder + OptiML-XOptimization=OptiML-X, Hardware=A100 80GB2026.02 | 9.1 | 0.87 | 46.37 | 27.91 | 58.24 | 1,183 | 4,368 | 2,172 |