May just be me, but article does seem a little sensationalist (as the website is geared toward machine learning on a Mac).
Can someone with more knowledge on the field elaborate on if this really is a breakthrough/a 'stunning' technological achievement?
I haven't much time to look at CoreML, but I do know a little bit how it's done before this new framework.
For instance, Inception v3 model was 'rewritten' to use in MetalPerformanceShaders(part of Metal framework on iOS). Since model provides parameters, you do know what 'layers' need to be used to implement inference network based on Inceptionv3. And this is quite difficult to do, looking at the code.
But here, bam - drag and drop, auto-generated classes and code, and you're ready to go, I guess. If it's working like this, I'm quite impressed.
Is this basically a path from model code that runs on an nvidia GPU like most people currently do, to a model that runs on iOS taking advantage of Metal?
No technological "breakthrough". It's more about availability and ease of use. They want to enable app developers to use ML on devices/in apps.