We have successfully used a virtual machine for building our applications with nvcc prior to CUDA 6. During the installation of CUDA toolkit, we chose to install driver and toolkit, which worked fine despite the fact that, of course, no NVIDIA GPUs are present on a virtual machine.
When we tried to do the same with CUDA 6, the driver installation failed. The toolkit itself installed correctly, but gave a warning that the installation is not fully functional as the driver is not installed. NVCC seems to be working fine though and produces binaries for devices with CC 1.0 - 5.0. What are the exact consequences of installing toolkit without a driver? What exactly will not be working? Why is a driver required to COMPILE CUDA (not to run it)?