We do know, or Xbox have misdescribed their solution. They used the word 'asymmetric'. Unified memory is not 'asymmetric', it is a technical data term in the same way : isochronous, asynchronous, bi-directional, half-duplex, full-duplex, etc all are.
GPU memory for efficiency reasons requires data to be (mostly) packed and/or packed and interleaved. Padding 192bits to 320bits just for unified access (at reduced speed for N number of cycles) doesn't sound deterministic or efficient, when a system that is already setup for client/server expects small GPU instruction transmissions in the form of client request, I'm now 99% certain from our discussion that the XsX does have to copy that data.
If you can't prove it isn't the normal way I have described, then (IMHO) it would seem fair – against the known history of GPU hardware and APIs – to agree that what I am inferring is the most probably solution, based on the information we have been given. And that would place the PS5 bandwidth in real workloads ahead of the XsX - as my RAID analogy proves and could be easily tested.