Reproduce “KernelClaw: A Multi-Agent Framework for GPU Kernel Optimization with Dual-Memory Mechanisms” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Qitong Sun, Jun Han, Tianlin Li, Zhe Tang, Sheng Chen, Fei Yang, Aishan Liu, Xianglong Liu and 1 more

KernelClaw uses profiling and source-code evidence to retrieve reusable optimization guidance, while separate memories track cross-task methods and task-specific attempts. It iteratively repairs and refines kernels. On KernelBench Levels 1–3, audited average speedup reaches 3.92× over Torch Eager at Level 2, versus 2.98× for AdaExplore. The advantage persists across three levels, but is limited to this setup and checker.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.