I’d welcome correction, but my understanding is that NPUs are “native” neural weight processors that trade the flexibility of doing graphics for the optimizatio

I’d welcome correction, but my understanding is that NPUs are “native” neural weight processors that trade the flexibility of doing graphics for the optimization of not having a layer like CUDA between neural net and hardware.