Coulomb, Landau and Maximally Abelian Gauge Fixing in Lattice QCD with Multi-GPUs

TL;DR

This paper develops a GPU-accelerated framework for lattice gauge fixing in SU(3) theories, optimizing performance and precision, and demonstrates linear scaling and high efficiency on multi-GPU systems.

Contribution

It introduces a CUDA-based lattice gauge fixing code supporting Coulomb, Landau, and maximally Abelian gauges with mixed precision schemes and multi-GPU scalability.

Findings

Achieves up to 3.5 Teraflops performance on 16 GPUs.

Demonstrates linear scaling with GPU count.

Provides a mixed precision scheme balancing accuracy and speed.

Abstract

A lattice gauge theory framework for simulations on graphic processing units (GPUs) using NVIDIA's CUDA is presented. The code comprises template classes that take care of an optimal data pattern to ensure coalesced reading from device memory to achieve maximum performance. In this work we concentrate on applications for lattice gauge fixing in 3+1 dimensional SU(3) lattice gauge field theories. We employ the overrelaxation, stochastic relaxation and simulated annealing algorithms which are perfectly suited to be accelerated by highly parallel architectures like GPUs. The applications support the Coulomb, Landau and maximally Abelian gauges. Moreover, we explore the evolution of the numerical accuracy of the SU(3) valued degrees of freedom over the runtime of the algorithms in single (SP) and double precision (DP). Therefrom we draw conclusions on the reliability of SP and DP simulations and suggest a mixed precision scheme that performs the critical parts of the algorithm in full DP while retaining 80-90% of the SP performance. Finally, multi-GPUs are adopted to overcome the memory constraint of single GPUs. A communicator class which hides the MPI data exchange at the boundaries of the lattice domains, via the low bandwidth PCI-Bus, effectively behind calculations in the inner part of the domain is presented. Linear scaling using 16 NVIDIA Tesla C2070 devices and a maximum performance of 3.5 Teraflops on lattices of size down to 64^3 x 256 is demonstrated.