Computational fluid dynamics （CFD） is a method to simulate the flow and heat transfer of fluids by numerically solving control equations. In civil aircraft design, CFD is widely used in wing design optimization, wind tunnel test validation, overall aircraft aerodynamic layout optimization, and fuel consumption evaluation. These problems involve complex turbulence, reactive flow, and multiphase flow, and require a large amount of computational resources due to large-scale grid calculations and a large number of state calculations. In order to improve computational efficiency, a GPU-based distributed parallel computing method is proposed. This method utilizes compute unified device architecture（CUDA） and message passing interface（MPI） technologies to perform parallel calculations on GPUs （graphic processing unit） and use MPI for communication between multiple GPUs. The method achieves parallelization of computational tasks and data transfer, and has been optimized for multi-stream parallelization and non-blocking communication, as well as load balancing between GPUs. The method is applied to the typical CFD case of supersonic plate flow, and compared with CPU serial computing, it achieves a speedup of 204 times using a single GPU, nearly 600 times using 4 GPUs, and more than 900 times using 8 GPUs on two nodes. It can be seen that this method has good parallel efficiency and computational performance, to some extent addressing the computational resource requirements of CFD applications.