A performance comparison of different graphics processing units running direct N-body simulations

Abstract

Hybrid computational architectures based on the joint power of Central Processing Units and Graphic Processing Units (GPUs) are becoming popular and powerful hardware tools for a wide range of simulations in biology, chemistry, engineering, physics, etc.. In this paper we present a comparison of performance of various GPUs available on market when applied to the numerical integration of the classic, gravitational, N-body problem. To do this, we developed an OpenCL version of the parallel code (HiGPUs) used for these tests, because this version is the only apt to work on GPUs of different makes. The main general result is that we confirm the reliability, speed and cheapness of GPUs when applied to the examined kind of problems (i.e. when the forces to evaluate are dependent on the mutual distances, as it happens in gravitational physics and molecular dynamics). More specifically, we find that also the cheap GPUs built to be employed just for gaming applications are very performant in terms of computing speed also in scientific applications and, although with some limitations in central memory and in bandwidth, can be a good choice to implement a machine for scientific use at a very good performance to cost ratio.

Sapporo2: a versatile direct N-body library

“…However, it also requires more operations than a fourth order integrator, something which is discussed in detail in [5]. Previous work [15,19,20] indicates that double-single accuracy is sufficient for a sixth order integrator. However, to give the user the choice we implemented both a double-single and a double precision version of this method.…”

“…This code is able to run on multiple GPUs and supports up to eighth order accuracy. In [19,20], the authors introduce the HiGPUs N -body code. This standalone code contains a sixth order integrator, and supports CUDA, OpenCL and IEEE-754 double precision accuracy.…”

Practical Implementation of Prestack Kirchhoff Time Migration on a General Purpose Graphics Processing Unit

“…Potentially, GPUs can achieve hundreds or even thousands of GFLOPS, whereas general CPUs are only capable of dozens of GFLOPS at present. NVIDIA's computed unified device architecture (CUDA) provides a C-like programming model for exploiting the massively parallel processing power of NVIDIA's GPU (NVIDIA 2013), and it is now employed widely for many parallel computation applications (Lu et al 2013, Capuzzo-Dolcetta and Spera 2013, Westphal et al 2014. Some studies have also used NVIDIA GPUs to accelerate PSTM.…”