Abstract: Sort-first architecture is often used to build high-performance parallel graphics rendering systems. Data-centralized sort-first systems based on immediate-mode deeply reply on network bandwidth, and thus bandwidth and belonging computing are likely to become system bottlenecks. An architecture of data distributed parallel rendering system based on retained-mode is presented. Geometry data is distributed to render nodes and adjusted between them when viewpoint is changed. Frame-to-frame coherence is efficiently utilized to lower overhead. Cell structure and a quick model space dividing algorithm are used to control parallel granularity. The experimental results show that high resolution display and parallel speedup can be achieved with relatively lower overhead.