A fast double precision CFD code using CUDA

Abstract

We describe a second order double precision finite volume Boussinesqcode designed to run on the CUDA architecture. We perform detailedvalidation of the code on a variety of Rayleigh-B´nard convectionproblems and show second order convergence. We obtain matching resultswith a Fortran code running on an eight-core CPU. The CUDA-acceleratedcode performs approximately eight times faster than the Fortran codeon identical problems. As a result, we are able to run a simulationwith a grid of size 3842 × 192 at 1.6 seconds per time step on amachine with a single GPU.