Abstract: By organizing the control mesh of subdivision in texture memory so that irregularities occur strictly inside independently refinable fragment meshes, all major features of subdivision algorithms can be realized in the framework of highly parallel stream processing. Our implementation of Catmull-Clark subdivision as a GPU kernel in programmable graphics hardware can model features like semi-smooth creases and global boundaries; and a simplified version achieves near-realtime depth-five re-evaluation of moderate-sized subdivision meshes. The approach is easily adapted to other refinement patterns, such as Loop, Doo-Sabin or sqrt(3) and it allows for postprocessing with additional shaders.