I found this on reddit which I reluctantly to cite it here [1], anyway the comments and the findings were as vague as Apple claiming they beat Nvidia RTX 3090 GPU with that fancy chart.
Regardless, all Apple current lineups, incl. Macbooks, Mac mini, Mac Studio Max come with 16-core Neural Engine, and the Ultra comes with 32-core Neural Engine.
What does it actually do despite all the marketing claims that none other than BS vague stuffs that only accessible to Apple proprietary apps, Finder, FaceTime, Final Cut Pro...
And from the schematic diagram of Apple M series SoC, the Neural Engine used significant space of the SoC.
Does Pytorch and other ML frameworks actually utilize that 16/32-core ?
[1] https://www.reddit.com/r/apple/comments/122iqf4/everything_we_actually_know_about_the_apple/
The new A17 Pro 3nm chip comes with hardware Ray Tracing, so we will expect the next gen Apple M3 series GPU performance boost. Hopefully, the gap with discrete GPU get closer.