Scaling SU(2) to 1000 GPUs using HiRep

TL;DR

This paper discusses the development of HiRep, a flexible lattice field theory simulation tool, and reports progress on scaling it efficiently to 1000 GPUs, especially on AMD hardware.

Contribution

It introduces HiRep's GPU port and demonstrates significant scaling improvements for large-scale simulations on AMD GPUs.

Findings

Successful GPU porting of HiRep features

Scaling to 1000 GPUs achieved

Enhanced capabilities for BSM physics simulations

Abstract

HiRep allows flexible simulations of higher representations of Wilson Fermions with various actions and gauge groups and a range of inverters and integrators. This is particularly important for enabling evaluations of observables relevant to phenomenological inputs for Beyond-the-Standard-Model physics from lattice field theory. We present progress on the GPU porting of available features, especially in terms of scaling to large jobs on AMD GPUs.