Whole-block fused megakernels, graded on decode speedup over optimized PyTorch.
Rankings
best left — click a column for cells and audits
KernelBench Megabest decode speedup vs optimized-PyTorch baseline over valid (correct + audited-clean) cells · H100 PCIe
KernelBench CUDAmean peak fraction of roofline over full deck (fails = 0) · RTX PRO 6000 (CUDA-only deck)
KernelBench Hardmean peak fraction of roofline over full deck (fails = 0) · H100 PCIe
The decks
pick a GPU board — frozen decks, public harnesses, traces on Hugging Face
CUDA-only writing deck — Triton and kernel DSLs fail the language gate.
Six-op CUDA/Triton deck, roofline-graded, one unlimited agent session per cell.
Multicoming soon
NVLink collectives rewritten as kernels on 8×H100 SXM, graded on busbw.
Performance vs compute
Does the win just cost more tokens? Output tokens = the compute each model chose to spend. Models with clean token telemetry.
on the efficiency frontier (most performance per token) dominated (spent more, delivered less)
Method
01
Roofline, not speedup
Scores ground in hardware ceilings; baseline quirks can't move them.
02
Real agent harnesses
Claude Code, Codex, Cursor, Kimi, OpenCode, Grok — the tools labs actually ship.
04
Judge-assisted audit
Reward hacks and rubric leaks get flagged, published, and linked per cell.
Cite this benchmark suite
@misc{arledge2026kernelbenchcom,
title = {kernelbench.com: Agentic GPU Kernel Benchmark Results and Run Artifacts},
author = {Arledge, Elliot},
year = {2026},
howpublished = {\url{https://kernelbench.com}},
note = {Website, benchmark results, transcript viewers, and citation index}
}
@misc{arledge2026hard,
title = {Hard: Agentic CUDA Kernel Result Suite},
author = {Arledge, Elliot},
year = {2026},
howpublished = {\url{https://github.com/Infatoshi/kernelbench.com/tree/master/benchmarks/hard}},
note = {CUDA benchmark suite, harness, results, and annotations}
}
@misc{arledge2026mega,
title = {Mega: Agentic GPU Megakernel Result Suite},
author = {Arledge, Elliot},
year = {2026},
howpublished = {\url{https://github.com/Infatoshi/kernelbench.com/tree/master/benchmarks/mega}},
note = {Megakernel benchmark suite, sandboxed harness, and result artifacts}
}
@misc{arledge2026hardtraces,
title = {KernelBench-Hard Agent Traces},
author = {Arledge, Elliot},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/Infatoshi/kernelbench-hard-traces}},
note = {Per-run agent transcripts: messages, tool calls, reasoning}
}
@misc{arledge2026megatraces,
title = {KernelBench-Mega Agent Traces},
author = {Arledge, Elliot},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/Infatoshi/kernelbench-mega-traces}},
note = {Per-run agent transcripts for the megakernel suite}
}