NNUE is weird. No one outside the chess/shogi community talks about it because they seem to have a very case of strong not-invented-here syndrome. It's hand-optimised CPU code that doesn't (yet) run on the GPU (which is why it is "faster").

To be fair, they do want to embed it into a consumer friendly application, and the integration for embedding TF or something that can run pytorch models on a GPU without python is non-trivial.

If someone ported it to a GPU, it almost certainly would be faster. See http://www.talkchess.com/forum3/viewtopic.php?f=7&t=76986

There is a PyTorch port available, but no benchmarks unfortunatly. It does seem to be fairly widely used for training though, which is indicative of the speed gains available.