I don't disagree with your conclusion but the comparison of max bandwidth between the two SoCs is not enough. Neither of them will use all of that bandwidth doing AI work because the GPU will be compute limited. That's why dedicated GPUs perform so significantly better without having significantly higher bandwidth.
“Comparable” is maybe true if we are talking about single core performance, but for memory bandwidth, the M1 Max is about 8 times faster. Wider bus, lower latency, not even close.
- "unified" ram makes all the system ram available as VRAM - dedicated ai coaccelerator thingy
Both of these reasons allow the apple silicon chips to crush conventional cpus in these kind of AI model workload stuffs
No idea about what the windows arm stuff is capable of. I know they use Qualcomm snapdragon chips though.