Thank you so much for this video. One thing to remember is that the NPU is really meant for lightweight inferencing (not training). It's designed for low-latency output rather than high-bandwidth. Also, CPU's with no integrated GPUs (and perhaps with cheap, low-end discrete graphics) would benefit greatly from having the NPU. Intel made a wise choice by integrating the NPU into the SoC rather than the GPU.
Surprisingly, we can't choose. That's why I didn't show it. Geekbench ML only uses DirectML with ONNX and we can't select which processor to use. The score is 3889 if you're curious.
who even cares about so-called "AI"? it's so over-done already. can't they just make a usable and comfortable OS? why do they always have to do some re-inventing and over-engineering for the simplest things, it's just pathetic.