What they're using here is still mostly off the shelf silicon with some tweaks. If they got enough volume, they probably could go for an all integrated APU with unified memory that could keep the GPU fed, but that'd be a very expensive and new thing to develop.

I hope that if this is a success, they'll have the numbers to justify a Strix-Halo like APU with a smaller CPU but keeping the big GPU for the next generation of the device.