Abstract: In order to keep the GPU accelerated Monte Carlo code to be able to handle 3D geometry and continuous energy point cross section, the method of geometry treatment acceleration by GPUs was proposed. The fission neutrons were organized into a neutron vector, and the geometry part of the Monte Carlo code was transplanted to GPUs. To reduce the negative impact of data communication on the performance of the accelerated code, CUDA streams were applied to design the asynchronous parallel algorithm. Two benchmarks including the fast reactor facility and the 17×17 PWR assembly were used for performance test. The results are satisfying and demonstrate that the speedup factor is close to the theoretical one for the local acceleration method.