Reproduce “Revealing Floating-Point Accumulation Orders in Software/Hardware Implementations” 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
Peichen Xie, Yanjie Gao, Yang Wang, Jilong Xue
FPRev is a black-box diagnostic that infers how software or hardware implementations group additions. It supplies carefully chosen inputs and uses observed outputs to reconstruct accumulation trees, assuming the order is fixed. In tests, half-precision PyTorch matrix multiplication through cuBLAS had 5-, 9-, and 17-way trees on V100, A100, and H100 GPUs, respectively; these findings do not establish application-level errors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.