Efficient implementation of the overlap operator on multi-GPUs

TL;DR

This paper presents efficient methods for implementing the overlap operator in lattice QCD calculations on multi-GPU systems, demonstrating significant performance advantages over CPU clusters.

Contribution

The paper introduces optimized implementation techniques for the overlap operator on multi-GPU architectures, addressing memory and parallelization challenges.

Findings

GPU clusters outperform CPU clusters by a factor of 20-30 in core count for similar performance

Efficient multi-GPU implementation reduces computational resources needed for lattice QCD

Demonstrates scalability of the overlap operator on GPU clusters

Abstract

Lattice QCD calculations were one of the first applications to show the potential of GPUs in the area of high performance computing. Our interest is to find ways to effectively use GPUs for lattice calculations using the overlap operator. The large memory footprint of these codes requires the use of multiple GPUs in parallel. In this paper we show the methods we used to implement this operator efficiently. We run our codes both on a GPU cluster and a CPU cluster with similar interconnects. We find that to match performance the CPU cluster requires 20-30 times more CPU cores than GPUs.