KernelBench Mega — H100 Best Decode Speedup
Code · 2026-06-18
KernelBench Mega best valid decode speedup on the H100 PCIe board. The live headline metric uses the Kimi-Linear W4A16 decode problem and excludes the sparsely attempted RL Grid PPO cell from the performance aggregate.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.8 | 15.5 |
| Kimi K3 | 14.8 |
| GPT-5.5 | 5.6 |
| Composer 2.5 | 1.8 |
| DeepSeek-V4-Pro | 1.4 |