This paper makes photon mapping algorithm capable of entirely running on latest GPUs by modifying its process in CUDA and harnesses the massive parallel computing power of GPU to accelerate the implementation of photon mapping. The implementation of photon mapping in CUDA includes the construction of the photon maps and estimation of the radiance. To accelerate the search of photons in photon map, this paper uses kd-tree to store photons and search them by KNN. The rendering speed is nearly 10 times by CPU under the test environment.