Whatever, but I think we can assume some serious data shifting between mobo/system resources via the taxed & maxed pcie bus & cpu. Clearly, only a fraction of the game can reside in gpu cache space.
Which is what vegaS killer feature of dedicated nvme on the separate Fabric bus is all about.
"Jul 30, 2017 - The Radeon Pro WX 9100 uses 16GB of High-Bandwidth Cache based on the ... The massive high-speed memory of the Radeon Pro SSG combined with the ..."
www.amd.com/en-us/press-releases/Pages/radeon-pro-wx-9100-2017jul30.aspx
(we can deduce they are using 4x nvmeS in raid 0 from the claimed bandwidth (also that they are 500GB panasonic 960 pro ssds - what else?
)).
Tho activated only on expensive vega ssd pro atm, nvme ports is supposed to be a standard vega feature/spec.
I would love to activate it on the cheap on lesser vega models (256GB samsung oem ssdS are pretty cheap and fast).
So to be clear, vast amounts of gpu data can swap in and out of gpu ram, independently of the pci bus, and 8GB can be made to seem like 16-24GB (amd claim) of gpu cache by hbcc intelligently pooling other HB resources like this and system memory.
This, and memory multiplying HBCC goodies, make me think vega is far more future proof than a bit of superficial Nvidia gaming muscle atm in the present gaming ecosystem.