Finite difference numerical simulations of acoustic fields with MPI and GPUS

The finite difference time domain (FDTD) method is widely implemented for numerical simulations of acoustic fields. When the size of the model is too large, the amount of the grid points in the computational region is huge, the simulation will cost a lot of calculation time when using CPU in a serial program. The paralleled method as MPI is not so convenient to be applied in the practical engineering field because a cluster computer with large volume is needed. In this paper, we use the graphic processing unit (GPU) to solve this problem. First, we use GPU to accelerate the numerical simulation of 3D eight-order FDTD for the isotropic medium model with PML absorb artificial edge and get a 50-time speedup. Then we accelerate a large 3-D model in horizontally layered isotropic medium and obtain a 40-time speedup with multi-GPUs and MPI method.