That makes sense for GPU-compute-in-the-cloud applications where you can pick the hardware and don't care if other hardware doesn't work, but some people want to write programs that use GPGPU that can run on anyone's machine.