Looks cool!
nvtop can actually support TPUs too via https://github.com/rdyro/libtpuinfo/ https://github.com/Syllo/nvtop/blob/76890233d759199f50ad3bdb...
Renaming fopen64 to intercept library calls feels like a brittle hack masquerading as "sandboxing." Why not just upstream this hardware support to nvtop instead of fragmenting the ecosystem?
"NPU" seems to refer to trainium only?
Look into all-smi https://github.com/lablup/all-smi It supports all GPUs thinkable including Apple Silicon and many AI accelerator cards.