Can SF use NPUs or CNNs?

Sort:
Buoyant27-1

Apple has 38 TOPS on M4 NPU, Intel has roughly 40 TOPS on Lunar Lake, and AMD Ryzen AI 300 series offers up to 50 peak TOPS. Stockfish (using NNUE) generally runs slower than Stockfish 11 (the last version before NNUE). Could this NPU benefit stockfish by increasing its speed?

In addition, I always thought with advancing hardware why doesn't Stockfish switch from it's current fully connected neural network (the NNUE), to a convolutional neural network (CNN), similar to networks used in AlphaZero and Leela Chess Zero?