In view of the problems and challenges in pre-stack reverse-time depth migration (RTM) for mass seismic data,the paper focuses on introducing GPU-RDMA and GPU-Stream,which are used to solve the large calculation cost and the mass memory demand of RTM on the CPU/GPU heterogeneous platform,and develops the RTM suitable for large-scale distributed parallel computing in GPU clusters.This technology can be used to deal with the high density seismic data,and the results of actual seismic data application show that the computing efficiency is up to 15~40 times.