2012 Freeman Scholar Lecture: Computational Fluid Dynamics on Graphics Processing Units

This paper discusses the various issues of using graphics processing units (GPU) for computing fluid flows. GPUs, used primarily for processing graphics functions in a computer, are massively parallel multicore processors, which can also perform scientific computations in a data parallel mode. In the past ten years, GPUs have become quite powerful and have challenged the central processing units (CPUs) in their price and performance characteristics. However, in order to fully benefit from the GPUs' performance, the numerical algorithms must be made data parallel and converge rapidly. In addition, the hardware features of the GPUs require that the memory access be managed carefully in order to not suffer from the high latency. Fully explicit algorithms for Euler and Navier-Stokes equations and the lattice Boltzmann method for mesoscopic flows have been widely incorporated on the GPUs, with significant speed-up over a scalar algorithm. However, more complex algorithms with implicit formulations and unstructured grids require innovative thinking in data access and management. This article reviews the literature on linear solvers and computational fluid dynamics (CFD) algorithms on GPUs, including the author's own research on simulations of fluid flows using GPUs.