A computing system includes a graphics processing unit (GPU) and main processing circuitry to execute computer program instructions forming a hypervisor, a control virtual machine (VM) and a specialized rendering VM for graphics processing. An application program of a guest VM generates graphics commands and data according to a graphics API such as Direct 3D. The rendering VM includes a graphics driver native to the GPU and is assigned pass-through access to the GPU by the control VM. The rendering VM receives the graphics information from the application program via an inter-VM communication channel, and it uses the graphics driver to control the GPU to perform graphics rendering. The use of the rendering VM enables native graphics performance to be achieved without constraining the control VM to use a compatible operating system. The technique is generally applicable to virtualization of hardware resources by specialized VMs.