Don't let that CPU sit idle: Hardware-aware heterogeneous general matrix multiplication (GEMM)

Date

2026

Authors

Warawa, Johnathan
Chester, Sean

Journal Title

Journal ISSN

Volume Title

Publisher

University of Victoria

Abstract

Matrix multiplication is a fundamental operation used to train neural networks for machine learning. GPUs are well-optimized for several stages of this operation and are thus used to accelerate the work, however, GPUs must be "hosted" by CPUs that remain underutilized while the GPU works, burning cycles that could be put to use by a more sophisticated, heterogeneous algorithm that makes use of both the GPU and CPU at the same time. In this project, we increase the speed of these operations beyond what a single processor could accomplish by developing a heterogeneous algorithm which efficiently divides and interleaves these operations.

Description

Keywords

GPU, SIMD, heterogeneous computing, GEMM, linear algebra, tensor cores, Jamie Cassels Undergraduate Research Awards (JCURA)

Citation