DataKernelBench: leaderboard
DataKernelBench (2026) asks LLMs to speed up the 22 TPC-H queries on an H100 GPU by writing CUDA or Triton kernels into validated PyTorch query programs, with execution-guided repair. Overall speedup over torch.compile, crediting an incorrect or too-slow kernel at the baseline time, for the best of four framework and scope configurations per model; higher is better.
Metric: Overall speedup over TorchPlan with torch.compile (times faster, end-to-end run_query runtime; a query whose kernel is incorrect or under the minimum speedup is credited at the torch.compile runtime; 22 TPC-H queries at scale factor 10 on one NVIDIA H100, each translated into a validated PyTorch TorchPlan; the model rewrites the query core or the full query in CUDA or Triton with up to several rounds of execution-guided repair; each model's best of its four framework-level configurations, as the leaderboard reports). Source: arxiv.org. Saturation forecast: Around June 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 2.11 | #49 |
| 2 | Claude Sonnet 4.6 | 1.54 | #161 |
| 3 | Claude Opus 4.7 | 1.51 | #79 |
| 4 | Gemini 3.1 Pro (Preview) | 1.44 | #94 |
| 5 | Claude Haiku 4.5 | 1.3 | #568 |
| 6 | GPT-OSS-120B | 1.26 | #700 |
| 7 | Qwen 3.5 397B A17B | 1.26 | #246 |
| 8 | DeepSeek V4 Flash | 1.23 | #258 |
| 9 | MiniMax-M2.5 | 1.19 | #582 |
| 10 | Devstral 2 | 1.08 | #1003 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=datakernelbench · How It Works · Data refreshed daily, snapshot 2026-11-13.