high performance computing on graphics processing units: hgpu.org

It is studied about parallel algorithm of lattice Boltzmann method. The data’s arrangement, commutation and computational progress are redesigned in a marriage of message passing interface and general purpose graphic processing Units. On the single-GPU, novel techniques appearing in shader model 3.0 such as frame buffer object (FBO), multiple-channels-rendering and, rendering-to-textures are used to improve computational efficiency. On multiple-GPUs, MPI is used to extend available mesh size and accomplish parallel algorithm. Consequently, the problem of excessively enormous mesh such as the size of 1024*1024, which could not be calculated on single GPU, is resolved in this paper. Moreover, the computational time of the instance-velocity vector of incompressible fluid is merely 0.585 second/step, a speed which is about 5.0 times faster than that of a single CPU implementation.