TeraFLOP computing on a desktop PC with GPUs for 3D CFD

A very efficient implementation of a lattice Boltzmann (LB) kernel in 3D on a graphical processing unit using the compute unified device architecture interface developed by nVIDIA is presented. By exploiting the explicit parallelism offered by the graphics hardware, we obtain an efficiency gain of up to two orders of magnitude with respect to the computational performance of a PC. A non-trivial example shows the performance of the LB implementation, which is based on a D3Q13 model that is described in detail.