This paper proposed a SAH-KD tree construction method for ray tracing to solve the problem of low parallel degree and low efficiency in the existing algorithms. The algorithm first obtains the ordered index on the three latitudes by sorting the primitive bounding boxes according to the left value. Then, according to the node bounding box, we choose the dimension to be divided and generate the candidate partition points of each node under this dimension. Finally, the SAH value of each candidate point was calculated based on the spatial tree in GPU, and the minimum SAH value of each node was selected for partitioning. Four common scenarios were used for testing performance of the algorithm, and compared the efficient serial and parallel GPU algorithm. The results show that the proposed algorithm can achieve 4~6 and 1.3~1.5 times faster than the contrast method when generating the same quality KD tree. When the number of GPU cores fold increases, the speedup ratio reaches a near multiple.