The Projected Tetrahedra is a popular method in the field of tetrahedra database visualization. Tretrahedra must be sorted according to obstruction between them to achieve an accurate rendering image, but strong dependency among tetrahedra results in not only inefficient sorting, but also poor parallel execution. This paper proposes a tetrahedra sorting algorithm which is based on K-D tree spatial partitioning. The database in one leaf node are peeled into layers in natural order, and the tetrahedra in the same layer are unobstructed . The peeling of different leaf node is independent, and their sorted tetrahedra are organized together according to the obstruction between leaf nodes. The sorting efficiency has improved greatly through two-level parallelism and guarantees accurate sorting. The data structure can be implemented easily in a graphics processing unit (GPU). The experimental results show that the quick and accurate sorting based on K-D tree processed in GPU shortens the sorting time greatly.