Or trying to install the whole of CUDA on machines that do not even have a GPU (not Nvidia, not anything past the embedded)...

Yep, that's (also) what I meant ;)

Lean, efficient... Also sensible and trouble-less.