Abstract: This paper presents a GPU-based parallel algorithm for radar simulation acceleration.The approach uses kD-tree as its acceleration structure.We implement several traversal algorithms of kD-tree on CUDA,and then find through a comparison test that the Shortstack-kD is the most suitable structure.Experimental results show that Shortstack-kD is always efficient for different models,with stack length can adjustable to control the balance between performance and flexibility.In addition,an optimization technique based on CUDA is applied to data structure of the scene and rays,suitable for device memory accessing of GPU.