GPU Kernel Optimization on KernelBench Level 3 (Fast1, Fast1.2, Fast1.5, Fast2)
32.2Fast1daVinci-kernel-14B
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| daVinci-kernel-14Bbackbone=Qwen3-14B-Base2026.06 | 32.2 | 7.3 | 2.3 | 1.3 | |
| Dr. Kernel-14B (reproduce)reproduce=true2026.06 | 22.1 | 2.8 | 1.8 | 1.5 | |
| GPT-5*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 21 | 12 | 4 | 2 | |
| Claude-4.5-Sonnet*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 21 | 11 | 5 | 4 | |
| Dr. Kernel-8B*data_source=Dr. Kernel paper2026.06 | 10.8 | 1 | 0 | 0 | |
| daVinci-kernel-8Bbackbone=Qwen3-8B-Base2026.06 | 10.1 | 3 | 2 | 0.8 | |
| Dr. Kernel-14B*data_source=Dr. Kernel paper2026.06 | 8.8 | 1.2 | 0.2 | 0 | |
| AutoTriton*data_source=Dr. Kernel paper2026.06 | 7.5 | 0 | 0 | 0 | |
| Qwen3-Coder-A30BA3*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 7 | 1 | 0 | 0 | |
| Qwen3-8B*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 5.7 | 0.2 | 0 | 0 | |
| GLM-4.7*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 5 | 2 | 2 | 2 | |
| Qwen3-32B*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 3.5 | 0 | 0 | 0 | |
| Dr. Kernel-8B (reproduce)reproduce=true2026.06 | 3 | 0.5 | 0.5 | 0.5 | |
| Deepseek-V3.2-Thinking*mode=zero-shot, data_source=Dr. Kernel paper2026.06 | 2 | 1 | 0 | 0 | |
| Cold-Start-14B (reproduce)reproduce=true2026.06 | 2 | 0.8 | 0.3 | 0.3 | |
| Cold-Start-8B*data_source=Dr. Kernel paper2026.06 | 0.5 | 0 | 0 | 0 | |
| Cold-Start-8B (reproduce)reproduce=true2026.06 | 0.5 | 0.3 | 0.3 | 0 |