The Role of Floating Point Precision in Two- and Three-Dimensional High Rayleigh Bénard Convection Modeled on Fermi GPU

We have implemented a second-order finite difference method for two-dimensional and three-dimensional Rayleigh-Benard thermal convection, corresponding to convection in the Earth´s mantle, on a single Fermi GPU. These codes are written in C for CUDA, making heavy use of CUBLAS routines for efficiency, and achieve performance on the order of 535 GFLOP/s and 100 GFLOP/s in single-precision and 230 GLFOP/s and 70 GFLOP/s in double-precision. We explore the sensitivity of this model to word length, finding that global characteristics remain constant despite a change in precision. Specifically, we compare the divergence between singleand double-precision runs with exactly identical initial conditions to the divergence between double-precision runs whose initial conditions have been perturbed by Gaussian noise. Our finding is that large-scale quantitative behavior (Nusselt number, number of plumes, etc) does not vary among these samples. This observation suggests a saving in time and computing resources could be enjoyed by implementing certain problems in single-precision. This is also valuable to scientists using iterative methods, as convergence may be completely unaffected by change of precision before the last few iterations. A particular interest is developed in the context of young Earth mantle convection, where higher Rayleigh numbers require both a finer computational mesh and a shorter timestep to properly resolve dynamic, small-scale features-compounding time wasted by inefficient or overly conservative computational implementations.