GPUs: An Oasis in the Supercomputing Desert

TL;DR

This paper introduces the 500/500 metric to compare national supercomputing resources, highlights Australia's disadvantage, and demonstrates GPU-based solutions for lattice QCD computations in single precision.

Contribution

It proposes a new metric for assessing national supercomputing capabilities and evaluates GPU methods for reliable lattice QCD calculations in single precision.

Findings

Australia scores poorly on the 500/500 metric.

Commodity GPUs are cost-effective for certain HPC tasks.

Modified CGNR provides reliable convergence in single precision.

Abstract

A novel metric is introduced to compare the supercomputing resources available to academic researchers on a national basis. Data from the supercomputing Top 500 and the top 500 universities in the Academic Ranking of World Universities (ARWU) are combined to form the proposed "500/500" score for a given country. Australia scores poorly in the 500/500 metric when compared with other countries with a similar ARWU ranking, an indication that HPC-based researchers in Australia are at a relative disadvantage with respect to their overseas competitors. For HPC problems where single precision is sufficient, commodity GPUs provide a cost-effective means of quenching the computational thirst of otherwise parched Lattice practitioners traversing the Australian supercomputing desert. We explore some of the more difficult terrain in single precision territory, finding that BiCGStab is unreliable in single precision at large lattice sizes. We test the CGNE and CGNR forms of the conjugate gradient method on the normal equations. Both CGNE and a modified form of CGNR (with restarts) provide reliable convergence for quark propagator calculations in single precision.