GPU Kernel Code Generation and Optimization on KernelBench Level 1 (test)
19CorrectnessStitchCUDA-G
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| StitchCUDA-GHardware=PRO 6000 (BlackWell)2026.03 | 19 | 3.97 | 55 | |
| CUDAForgeHardware=PRO 6000 (BlackWell)2026.03 | 18 | 2.74 | 50 | |
| StitchCUDAHardware=PRO 6000 (BlackWell)2026.03 | 18 | 2.86 | 35 | |
| CUDAForgeHardware=H200 (Hooper)2026.03 | 18 | 3.18 | 45 | |
| StitchCUDA-GHardware=H200 (Hooper)2026.03 | 18 | 3.57 | 45 | |
| StitchCUDAHardware=H200 (Hooper)2026.03 | 18 | 3.54 | 55 | |
| StitchCUDA-QHardware=H200 (Hooper)2026.03 | 17 | 2.13 | 20 | |
| StitchCUDA-KHardware=PRO 6000 (BlackWell)2026.03 | 13 | 2.11 | 30 | |
| Claude-4-sonnetHardware=PRO 6000 (BlackWell)2026.03 | 12 | 1.02 | 25 | |
| Claude-4-sonnetHardware=H200 (Hooper)2026.03 | 12 | 1.04 | 30 | |
| StitchCUDA-KHardware=H200 (Hooper)2026.03 | 12 | 2.25 | 50 | |
| GPT5.2Hardware=PRO 6000 (BlackWell)2026.03 | 10 | 0.73 | 25 | |
| GPT5.2Hardware=H200 (Hooper)2026.03 | 10 | 0.6 | 20 | |
| Kevin32BHardware=PRO 6000 (BlackWell)2026.03 | 9 | 0.65 | 10 | |
| Kevin32BHardware=H200 (Hooper)2026.03 | 9 | 0.62 | 10 | |
| StitchCUDA-QHardware=PRO 6000 (BlackWell)2026.03 | 3 | 0.19 | 5 | |
| Qwen3-32BHardware=PRO 6000 (BlackWell)2026.03 | 2 | 0.02 | 5 | |
| Qwen3-32BHardware=H200 (Hooper)2026.03 | 2 | 0.21 | 5 |