Pseudopotential model of lattice Boltzmann method is partially non-local for pseudopotential calculation with coupling of lattices, which leads to synchronization of threads in parallel implementation process. Besides, it uses a large number of registers and much time of data access operations when access global memory in calculation process. They lead to low computational efficiency. In this paper, a multi-relaxation-time(MRT) 3D pseudopotential model with D3Q15 lattice is adopted as an example to investigate performance of parallel computing based on GPU. To address limitation of parallel computing of pseudo-potential model, efficiency of reading and writing of global memory is improved by using merge access method. To improve efficiency of grids retrieving data which are in boundary of lattice, a "Directional Transfer" algorithm is proposed. The role of computing resource configuration is investigated with different sizes of block, and optimal resource configuration scheme is obtained.