I will believe this when I see it. It’s totally possible that those capabilities are locked behind some private API or that there’s some weedsy hardware complication not mentioned that makes them non-viable for what we want to do with them.
They might recommend using CoreML to leverage them, though I imagine it will be available to Metal.
The whole point of CoreML is that your solution uses whatever hardware is available to you, including enlisting a heterogeneous set of units to conquer a large problem. Software written years ago would use the GPU matmul if deployed to a capable machine.
