Recently NVidia proposed their first cloud based GPU architecture called Kepler. One of the key features of Kepler GPU is that it allows multiple user processes to access the GPU {em concurrently}. We use this feature to design a cloud computing system that allow multiple users to share the computing power of a Kepler board anytime, anywhere over the internet. Our system improves the utilization of the Kepler GPU and lowers the cost in providing GPU cloud services. We conduct experiments to evaluate the overhead of our system, and preliminary results indicatse that our system provides convienent services with very little overhead.