GPU Kernel Generation and Optimization on KernelBench Level 2
100AccuracyKernel Foundry
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Kernel FoundryOptimizer=Qwen-3.5, Expert=AutoTriton2026.05 | 100 | 64 | 1.06 | 0.94 | |
| Kernel FoundryOptimizer=Claude-sonnet-4.5, Expert=AutoTriton2026.05 | 100 | 62 | 1.03 | 0.96 | |
| Kernel FoundryOptimizer=ChatGPT-5.4, Expert=AutoTriton2026.05 | 98 | 61 | 2.46 | 1.01 | |
| Kernel FoundryOptimizer=Deepseek-V3, Expert=AutoTriton2026.05 | 97 | 71 | 1.08 | 1.17 | |
| AutoTriton2026.05 | 94 | 58 | 0.99 | 1.01 | |
| AutoTritonevaluation protocol=w/o cheating2026.05 | 92 | 50 | 0.95 | 0.98 | |
| Kernel FoundryOptimizer=Deepseek-V3, Expert=KernelLLM2026.05 | 45 | 38 | 0.56 | 1.22 | |
| Kernel FoundryOptimizer=Deepseek-V3, Expert=No Expert2026.05 | 40 | 26 | 0.43 | 1.05 | |
| KernelLLM2026.05 | 32 | 31 | 0.45 | 1.38 | |
| KernelLLMevaluation protocol=w/o cheating2026.05 | 28 | 27 | 0.39 | 1.35 | |
| DeepSeek-V32026.05 | 23 | 17 | 0.26 | 1.09 |