Texture synthesis is important in producing computer-generated animation. In order to tackle the issue of poor efficiency of traditional serial point matching texture synthesis algorithm, we propose a parallel approach based on CUDA (Compute Unified Device Architecture) for texture synthesis. By reasonably arranging data transfer between CPU and GPU, and using GPU to do tedious and time-consuming computing, we are able to greatly improve the efficiency of texture synthesis. Testing the parallel algorithm on two systems with different level of Nvidia Graphics Cards, the experiment results indicate that the GPU parallel algorithm achieves a speed-up ratio up to 160.