Direct Numerical Simulations (DNS) of a turbulent plane jet, at Reynolds number are performed using a programmable commodity graphics processors units (GPUs) assisted solver. The mesh comprises more than one hundred million points and the numerical scheme uses both pseudo spectral methods and compact differences. Time advancement is done via a low storage Runge-Kutta method.
The accuracy and performance of the GPU assisted simulations and their relative advantages/drawbacks over CPU based computation are reported. The GPU computation achieved speedups up to 30% against CPU calculations. The relatively low gains obtained are attributed to the low parallelism of the spectral and compact difference schemes coupling encountered in the original DNS code.