Reproduce “EvoCG: Self-Evolving Precision Control for Faster Conjugate Gradient” 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

Jiaqi Zhai, Haozhi Han, Kaiyi Huang, Wenju Zhao, Yifeng Chen, Yunquan Zhang, Ting Cao, Yunxin Liu and 3 more

EvoCG combines a learned controller that selects FP64, FP32, or BF16 precision tile by tile during conjugate-gradient solves with Tile-CSC storage and execution designed for those choices. Across 47 SuiteSparse matrices on an NVIDIA A100, EvoCG-Mixed is 1.28× faster geometrically than its all-FP64 counterpart, with 11.4% more iterations and a 1.0108 ratio of separately checked true-residual norms, on this benchmark setup.

Summary by alphaXiv.

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