LRAMM -- Low precision approximates GEMM via RSVD

TL;DR

This paper introduces LRAMM, a high-performance approximate matrix multiplication method that combines mixed-precision quantization with RSVD and low-rank decomposition to improve efficiency while maintaining acceptable accuracy.

Contribution

LRAMM is a novel algorithm integrating mixed-precision quantization and RSVD, optimizing matrix multiplication speed with controlled error margins.

Findings

Significant speedup in matrix multiplication tasks.

Maintains accuracy within low-precision error bounds.

Effective use of low-rank decomposition for approximation.

Abstract

Matrix multiplication computation acceleration has been a research hotspot across various domains. Due to the characteristics of some applications, approximate matrix multiplication can achieve significant performance improvements without losing much precision. In this paper, we propose LRAMM - a high-performance matrix multiplication approximation algorithm that combines mixed-precision quantized matrix multiplication with RSVD techniques, further enhancing efficiency within the error range of low-precision matrix multiplication by utilizing matrix low-rank decomposition technology.