Reproduce “Mechanistic Critics for Sample-Efficient NPU Design-Space Exploration” 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

Yicheng He, Yuqi Xue, Tianyang Wu, Jian Huang, Bin Hu

CriticNPU guides NPU architecture selection for LLM deployments by screening configurations before simulation and using operator timings to diagnose bottlenecks and steer proposals. Across six workloads and 16 deployment settings per workload, it achieves best-or-tied-best latency in 74% of prefill settings and 87.5% of decode settings, with about 54 simulator calls per run. The evaluation uses NeuSIM simulations, not hardware.

Summary by alphaXiv.

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