GPU Kernel Code Generation and Optimization on KernelBench Level 3 (test)
100Correctness ScoreStitchCUDA
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| StitchCUDAHardware=PRO 6000 (BlackWell)2026.03 | 100 | 1.27 | 70 | |
| StitchCUDAHardware=H200 (Hooper)2026.03 | 90 | 1.5 | 70 | |
| StitchCUDA-KHardware=PRO 6000 (BlackWell)2026.03 | 70 | 0.42 | 0 | |
| CUDAForgeHardware=PRO 6000 (BlackWell)2026.03 | 60 | 0.75 | 30 | |
| StitchCUDA-QHardware=PRO 6000 (BlackWell)2026.03 | 60 | 0.4 | 10 | |
| StitchCUDA-GHardware=PRO 6000 (BlackWell)2026.03 | 60 | 0.99 | 50 | |
| CUDAForgeHardware=H200 (Hooper)2026.03 | 60 | 0.87 | 40 | |
| StitchCUDA-GHardware=H200 (Hooper)2026.03 | 60 | 1.01 | 50 | |
| StitchCUDA-KHardware=H200 (Hooper)2026.03 | 50 | 0.32 | 0 | |
| Kevin32BHardware=PRO 6000 (BlackWell)2026.03 | 40 | 0.22 | 0 | |
| Claude-4-sonnetHardware=PRO 6000 (BlackWell)2026.03 | 30 | 0.67 | 20 | |
| Claude-4-sonnetHardware=H200 (Hooper)2026.03 | 30 | 0.64 | 20 | |
| StitchCUDA-QHardware=H200 (Hooper)2026.03 | 30 | 0.24 | 10 | |
| GPT5.2Hardware=PRO 6000 (BlackWell)2026.03 | 20 | 0.47 | 10 | |
| GPT5.2Hardware=H200 (Hooper)2026.03 | 20 | 0.48 | 10 | |
| Kevin32BHardware=H200 (Hooper)2026.03 | 20 | 0.34 | 10 | |
| Qwen3-32BHardware=PRO 6000 (BlackWell)2026.03 | 0 | 0 | 0 | |
| Qwen3-32BHardware=H200 (Hooper)2026.03 | 0 | 0 | 0 |