GPU kernel optimization on KernelBench Level 1 (Fast1, Fast1.2, Fast1.5, Fast2)
37.2Fast1 ScoredaVinci-kernel-14B
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| daVinci-kernel-14Bbackbone=Qwen3-14B-Base2026.06 | 37.2 | 27.3 | 22.9 | 16.2 | |
| Dr. Kernel-14B (reproduce)reproduce=true2026.06 | 30.4 | 26 | 20.4 | 13.9 | |
| daVinci-kernel-8Bbackbone=Qwen3-8B-Base2026.06 | 26.1 | 20.4 | 16.8 | 13 | |
| Dr. Kernel-14B*data_source=Dr. Kernel paper2026.06 | 20.3 | 16.9 | 13.2 | 11.6 | |
| GPT-5*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 19.5 | 16.5 | 12.5 | 11 | |
| GLM-4.7*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 19.4 | 17.2 | 13.1 | 10.4 | |
| Dr. Kernel-8B (reproduce)reproduce=true2026.06 | 17.9 | 14.9 | 10.5 | 7.2 | |
| Dr. Kernel-8B*data_source=Dr. Kernel paper2026.06 | 15.9 | 12.8 | 10.9 | 8.4 | |
| Claude-4.5-Sonnet*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 15.5 | 13.5 | 11 | 8.5 | |
| Cold-Start-14B (reproduce)reproduce=true2026.06 | 13.6 | 11.6 | 8.5 | 6.4 | |
| Cold-Start-8B (reproduce)reproduce=true2026.06 | 11.4 | 9.3 | 7.4 | 4.9 | |
| Deepseek-V3.2-Thinking*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 7.5 | 5.5 | 4.5 | 4 | |
| Cold-Start-8B*data_source=Dr. Kernel paper2026.06 | 7.5 | 6.6 | 5 | 4.3 | |
| Qwen3-32B*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 6.1 | 4.9 | 4.3 | 4 | |
| Qwen3-Coder-A30BA3*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 6 | 5.2 | 5.1 | 3.8 | |
| Qwen3-8B*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 5.8 | 4.8 | 4.1 | 3.4 | |
| AutoTriton*data_source=Dr. Kernel paper2026.06 | 4.5 | 3.6 | 2.8 | 2.1 |