Agentic GPU kernel benchmark results

For business inquiries reach out to elliot@arledge.net.

Mega

Full GPU megakernel build (Kimi-Linear W4A16 decode + RL), frontier models across Blackwell / H100 / B200, scored as speedup over the reference. Contamination-audited; agent sandbox hides prior solutions.

problem
megakernel
models
11+
GPUs
3

Speedup over reference megakernel

Kimi linear-decode megakernel, per model. Pick a GPU.

Claude Fable 5
18.7x
Claude Opus 4.8
14.4x
GLM-5.2
11.1x
GPT-5.5
4.3x
Claude Sonnet 5
4.0x
MiniMax-M3
2.6x
Kimi K2.7-Code
2.6x
Composer 2.5 Fast
2.5x
Gemini 3.5 Flash
2.3x
DeepSeek V4 Pro
2.1x
GPT-5.6 Sol
1.1x
Grok 4.5
0.8x
Show data table
ModelRTX PRO 6000H100B200
Claude Opus 4.814.4x15.5x19.4x
Claude Fable 518.7xDNFDNF
GLM-5.211.1xDNF7.3x
GPT-5.54.3x5.6x9.4x
MiniMax-M32.6xDNF4.2x
Claude Sonnet 54.0xDNFDNF
Gemini 3.5 Flash2.3x2.7x2.5x
Kimi K2.7-Code2.6xDNFDNF
Composer 2.5 Fast2.5x1.8x1.2x
DeepSeek V4 Pro2.1x1.4x1.6x
GPT-5.6 Sol1.1xDNFDNF
Grok 4.50.8xDNFDNF
GitHub sourceHugging Face runs

Hard

Curated hard CUDA/Triton deck — Opus, Fable, Grok 4.5, GPT-5.6 Sol, GLM-5.2, and more — across RTX PRO 6000 Blackwell, H100 PCIe, and B200. Roofline-graded with per-run agent traces.

problems
6
models
13
GPUs
3+

Percent of hardware roofline

Average across solved problems, relative to the GPU's own roofline. Pick a GPU.

GPT-5.6 Sol
30%
GLM-5.2
26%
Claude Opus 4.8
24%
Claude Fable 5
23%
Claude Sonnet 5
23%
MiniMax-M3
22%
GPT-5.5
22%
Grok 4.5
21%
DeepSeek V4 Pro
19%
Kimi K2.7-Code
17%
LongCat 2.0
16%
Composer 2.5 Fast
15%
Gemini 3.5 Flash
12%
Show data table
ModelRTX PRO 6000H100B200
GPT-5.6 Sol30%DNFDNF
GLM-5.226%11%10%
Claude Opus 4.824%21%10%
Claude Fable 523%24%14%
Claude Sonnet 523%DNFDNF
MiniMax-M322%13%4%
GPT-5.522%14%8%
Grok 4.521%DNFDNF
DeepSeek V4 Pro19%3%4%
Kimi K2.7-Code17%10%6%
LongCat 2.016%10%DNF
Composer 2.5 Fast15%9%5%
Gemini 3.5 Flash12%6%5%
GitHub sourceHugging Face runs

Performance vs compute

Does the win just cost more tokens? Output tokens = the compute each model chose to spend. Models with clean token telemetry.

0.0x2.0x4.0x6.0x8.0x10.0x12.0x14.0x16.0x18.0x20.0x0k100k200k300k400k500k600koutput tokens spentGeminiGPT-5.6 SolComposerGPT-5.6 SolDeepSeekGPT-5.5GLM-5.2MiniMaxSonnet 5Opus 4.8Fable 5

on the efficiency frontier (most performance per token) dominated (spent more, delivered less)

Design principles

  • •Percent of theoretical maximum, not speedup over PyTorch. Scores are grounded in hardware ceilings instead of baseline quirks.
  • •Modern coding-agent harnesses. Runs use Claude Code, Codex CLI, Cursor, Gemini CLI, Kimi CLI, OpenCode, Grok, and MiniMax where those are the natural interfaces.
  • •Public transcript viewers. Browse the run index or open the Hard agent traces on Hugging Face to inspect tool calls, reasoning, solution files, and timing.
  • •Judge-assisted audit. A judge model helps flag reward hacking, rubric leaks, and suspicious shortcuts for human review.

Citation

Cite this benchmark suite

If you use these results, cite the website for the public benchmark view, the relevant benchmark repository for problem definitions and harness code, and the Hugging Face dataset for the run transcripts. This is a results and artifact site, not Stanford's original KernelBench benchmark.

WebsiteWebsite repositoryMega repositoryHard repositoryMega HF tracesHard HF traces
@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}
}