We would like to make archspec aware of GPUs, probably starting with NVIDIA GPUs (since they're most prevalent in the HPC community currently).

Before diving in, I would like to collect some feedback/ideas in this issue.

For example, is running the nvidia-smi the best way to obtain information about the GPU(s) available?
How do we deal with having multiple different GPUs in a single system (do people do that?)?

Should we make archspec aware of which CUDA compute capabilities are supported by a GPU, and what the corresponding options for the nvcc compiler are?