GPU Kernel Optimization on KernelBench Level 3
5.3Average SpeedupKernelPro
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| KernelProLLM=Sonnet 4.6, GPU=A100, Metric Type=GeoMean (G), Tasks=50 tasks2026.06 | 5.3 | — | — | — | — | — | — | — | — | |
| KERNELBLASTERHardware Setup=L40S, Agent's Models=GPT-4.1 and GPT-5.0, Hyperparameter=10 iterations, 10 rollout steps per iteration2026.02 | 1.749 | 67 | 1.502 | 1.894 | 0.529 | 2.681 | 75 | 25 | — | |
| KernelBlasterLLM=GPT-4.1/5.0, GPU=H100, Metric Type=GeoMean (G), Tasks=L3 subset (size unspecified)2026.06 | 1.5 | — | — | — | — | — | — | — | — | |
| StitchCUDALLM=Qwen3-32B (Coder), GPT-5.2 (Planner/Verifier), GPU=H200, Metric Type=ArithMean (A), Tasks=10 test tasks, Training Protocol=In-distribution (trained on 80% of KernelBench)2026.06 | 1.5 | — | — | — | — | — | — | — | — | |
| KERNELBLASTERHardware Setup=H100, Agent's Models=GPT-4.1 and GPT-5.0, Hyperparameter=10 iterations, 10 rollout steps per iteration2026.02 | 1.336 | 67 | 1.11 | 1.333 | 0.375 | 2.302 | 75 | 25 | — | |
| CudaForgeLLM=o3, GPU=RTX 6000, Metric Type=ArithMean (A), Tasks=50 tasks2026.06 | 1.28 | — | — | — | — | — | — | — | — | |
| KernelBlasterHardware=A1002026.06 | — | — | 1.5 | — | — | — | — | — | — | |
| KernelProHardware=A1002026.06 | — | — | 5.3 | — | — | — | — | — | 50 |