GPU Accelerated Hall Magnetohydrodynamics

We describe our experience porting a new Hall MHD solver to an NVIDIA GPU. The algorithm is a finite volume scheme on a uniform grid. Numerical fluxes are computed using a second-order MUSCL-Hancock scheme with HLL Riemann solver. Strang dimensional splitting is used to achieve second-order temporal accuracy. Magnetic field divergence errors are controlled via Dedner's hyperbolic divergence cleaning method. We tested our code on a GTX 480 (Fermi architecture) and achieved a double precision speedup (relative to the sequential algorithm running on a single Nehalem core) of 58 for a 1024x1024 mesh. We discuss potential challenges involved in using CPU-GPU clusters for large scale geospace physics applications.