It'll probably behave as it does now if you try using unified memory in Metal (i.e., you will get an error). It's currently only valid in iOS/iPad/iEtc devices.

Have a look here: https://developer.apple.com/documentation/metal/mtlstoragemo...

I guess there will be a compile-time flag that you can #ifdef to check if you are on a new Apple Silicon mac or not.

It looks like Apple will be going it's own way with GPU, as tight integration between GPU and CPU enables much better performance. Think vias rather than buses.

The higher end Macs should see some large on-chip (or part of chiplet array) GPU implementations.

It'll largely be transparent at the Metal API level.

Interestingly, from a business perspective this is mainly bad for AMD.