Abstract: GRAPES(globalregional assimilation and prediction system) is a typical non-linear discrete system of the Earth’s atmosphere developed for numerical weather prediction. There are heavy computations involved in GRAPES. Researchers have recently paid a lot of attentions to the parallel acceleration of the GRAPES model by low-cost, low-power, and high-performance GPUs. In this paper, we implement the parallel acceleration for the GRAPES model in GPUs. But the experimental results show that its performance is not efficient as supposed. Therefore, based on this, we further propose some strategies for optimizing the system performance, including reducing the data transmission time, decreasing the amount of device memory loaded and stored equipment, and avoiding the branches of thread control flows. The experimental results show that the performance optimizations on GPUs improve GRAPES system performance effectively.