Because of low efficiency, seismic forward simulation of 3D wave equation by finite difference method cannot be applied to real large-scale product. To deal with this problem, the authors present a parallel accelerating solution based on multi-GPUs heterogeneous cluster. By using CPU/GPU cooperation parallel mode and domain decomposition method, the authors carry out a multi-level parallel strategy of this algorithm. Furthermore, non-blocking MPI communications, storage optimization and parallel I/O mechanism using MPI-IO API are presented to optimize computation efficiency and memory usage. This parallel 3D forward algorithm can effectively support large-scale practical production.