AgentKernelArena - PyTorch-to-HIP: leaderboard
GPU kernel optimization by coding agents on AMD Instinct MI300X from AgentKernelArena (Younesian et al., AMD, 2026): agents write HIP kernels from scratch for 24 PyTorch modules and beat PyTorch eager execution, working in an isolated workspace with compile, correctness and performance gates. Mean per-task speedup (a ratio, unbounded above), with failed compilation or correctness counted as 0, averaged over three runs; higher is better.
Metric: Mean speedup (x; arithmetic mean over the 24 PyTorch-to-HIP tasks of the per-task speedup of the agent-written HIP kernel over PyTorch eager execution, counting 0 for a task whose kernel fails compilation or correctness; per-task mean of 3 runs; agent in an isolated workspace, 3600 s timeout, up to 3 kernel versions; AMD Instinct MI300X). Source: arxiv.org. Saturation forecast: Not forecast. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 (High) | 6.89 | #100 (Claude Opus 4.6) |
| 2 | Claude Opus 4.6 | 6.7 | #100 |
| 3 | Claude Opus 4.7 (High) | 6.65 | #75 (Claude Opus 4.7) |
| 4 | Claude Sonnet 4.6 | 5.3 | #158 |
| 5 | GPT-5.3 Codex | 5.2 | #114 |
| 6 | GPT-5.4 (High) | 3.85 | #121 (GPT-5.4) |
| 7 | GPT-5.3 Codex (High) | 3.74 | #114 (GPT-5.3 Codex) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=agentkernelarena-pytorch-to-hip · How It Works · Data refreshed daily, snapshot 2026-11-12.