In this paper, the authors illustrate the implementation process of RTM and analyze the stability condition and dispersion relation of finite difference (FD) method. For the problem of intensive computation cost, the authors use the Graphic Processing Unit (GPU) architecture to realize RTM and get an order of magnitude higher speedup ratio compared to the traditional CPU architecture. Compared to the one way wave equation migration methods, RTM does not have the imaging dip limit and the imaging effect is significantly improved. The test on vertical fault and BP models proves this conclusion.