A system and method is provided for performing computations on a virtual machine without a special hardware computation unit, such as a discrete graphics processing unit (GPU). The described method uses a computation module to intercept requests from a user application executing in a virtual machine on a first physical computer. The intercepted requests may include requests to configure GPU computation grids, start and finish accelerated code execution, and transfer data to and from the special computation unit. The computation module offloads accelerated code to a second physical computer having a physical special hardware unit (e.g., discrete GPU).