Abstract: In this paper, we develop a tiled-based ray tracing technique. This algorithm divides the acceleration data structure into tiles or tree-lets, which can fit onto GPU's on-chip memory, and traverses each part with corresponding rays. The coherence of rays and data structure can significantly reduce the demand for GPU bandwidth. Experiments prove that our algorithm improves the traversal efficiency of traversal of incoherent rays (e.g., diffuse rays) to a level that is comparable with that of highly coherent primary rays.