Abstract:Radar networking method is one of the most important methods in Severe Weather Automatic Nowcasting （SWAN） radar applications. Improving the efficiency of the radar networking method can improve the timeliness of the short term algorithm sequence, and also can make good use of high resolution radar data, which has important practical significance. With the development of GPU (graphics processing unit) general computing interface, GPU has also been regarded as a powerful computing resource rather than display device for rendering and images. Therefore, this study proposes a new GPU parallel radar networking method based on CUDA (compute unified device architecture). The method is designed in a hybrid architecture of CPU (central processing unit) and GPU, in which the CPU is for the decoding of radar data and scheduling the GPU parallel modules, and the GPU is for the parallel computing of large scale data. By analyzing the parallel overhead of the CUDA and the characteristics of the radar networking method, a scheme of GPU memory management optimization and data exchange process simplification is proposed and implemented, which effectively improves the efficiency of the method. The comparative test results show that, compared with the CPU parallel algorithm in SWAN, the GPU parallel networking method based on CUDA achieves a speed up ratio of 3.52 and 6.82, respectively, on the national puzzle tasks of 1 km and 500 m resolution. To sum up, the parallel networking method based on CUDA can not only improve the timeliness of the short term nowcasting methods sequence, but also provide technical support for the puzzle of higher resolution radar data.