Mixed-Precision Communication-Avoiding SGD for Generalized Linear Models on GPUs

DOI10.48550/arXiv.2606.18463MaRDI QIDQ7266589arXivFDO

Authors Aditya Devarakonda, Giulia Guidi, Irene Simo Munoz

Publication date 16 June 2026

Published in arXiv

Tldr This paper studies mixed-precision communication-avoiding SGD for generalized linear models on NVIDIA GPUs, decomposing rounding error into nine independent precision choices tied only to hardware unit roundoffs. On NERSC Perlmutter A100 GPUs, the method matches FP32 SGD loss within 0.5% while achieving 5.1-6.8x speedup.

Full work available at URL https://doi.org/10.48550/arXiv.2606.18463

arXiv classification

cs.DCcs.LGmath.NAstat.ML

This page was built for publication: Mixed-Precision Communication-Avoiding SGD for Generalized Linear Models on GPUs

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7266589)