[fho]

I guess the simplest approach to port NNUE to GPUs would be to run a complete instance per GPU thread (ie concurrent, not parallel evaluation).