Lattice-Boltzmann Navier-Stokes simulation on Graphic Processing Units

PABLO R. RINALDI; ENZO A. DARI; MARCELO J. VÉNERE; ALEJANDRO CLAUSSE

Asian Journal of Applied Sciences

Knowledgia Review

Año: 2011 vol. 4 p. 762 - 770

1996-3343

Lattice Boltzmann Methods (LBM) was one of the first simulation models which successfully run on Graphic Processing Units. Earlier LBM implementations using NVIDIA Compute Unified Device Architecture programming language required two steps (collision-propagation and exchange) to maximize memory bandwidth. This article presents a parallel single-step implementation of the Lattice Boltzmann method with fully coalesced memory access using shared-memory. The resulting code, running on low cost Personal Computer Graphic Processing Unit was able to process more than 800 million cells updates per second, approaching High Performance Computing clusters’ performances. Substantial reductions of the calculation rates were achieved, lowering down to 240 times the time required by a CPU to execute the same model. The code was tested on the numerical calculation of the flow in a two-dimensional channel with a sudden expansion. The precision of the results were validated against a proved finite-element solver.