In this paper, we consider a real-time MPEG-4 streaming architecture to facilitate remote visualization of large scale 3D models on thin clients, which denote most of the hand-held devices that have limited computing resources. MPEG-4 serves as a key component to handle the compression, transmission, and visualization of the high-end supercomputer rendered image sequence, allowing the synchronization of the data in both the terminal and the server. The MPEG-4 encoding speed is thus the bottleneck of the system, in particular, the motion estimation process takes more than half of the total encoding time. We propose a fast motion estimation algorithm that expedites the MPEG-4 encoding process. Our algorithm utilizes the 3D data available at the server and is able to directly calculate the motion vector on a block basis without having to employ the expensive MPEG motion searching procedure. In addition, our algorithm can be implemented on the Graphic Processor Units(GPUs) such that most of the motion estimation process can be done in parallel to the encoding process. Our preliminary results show that the proposed motion estimation is able to significantly speed up the encoding process while maintaining the encoding quality. 1 1