KernelBench: leaderboard
250 PyTorch workloads (operators, fusions, full architectures) where LLMs generate CUDA kernels. Measures average speedup over PyTorch baseline on Nvidia L40S. From Stanford Scaling Intelligence Lab.
Metric: Avg Speedup vs PyTorch. Source: scalingintelligence.stanford.edu. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 | 1 | #183 |
| 2 | Claude 3.5 Sonnet | 0.87 | #350 |
| 3 | GPT-4o | 0.79 | #343 |
| 4 | Llama 3.1 405B | 0.7 | #466 |
| 5 | Gemini 1.5 Flash | 0.64 | #618 |
| 6 | Llama 3.1 70B | 0.54 | #592 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=kernelbench · How It Works · Data refreshed daily, snapshot 2026-10-20.