One reason is to mix AMD and NVIDIA GPUs – I experimented with running X on the integrated Radeon device, while running CUDA on a Linux guest with GPU-passthrou

One reason is to mix AMD and NVIDIA GPUs – I experimented with running X on the integrated Radeon device, while running CUDA on a Linux guest with GPU-passthrough. Trying to have CUDA on the host system led to library conflicts.