It's debatable, I think. Cross gen will be around for a while IMHO. They may have hit their peak market penetration at <50M units, certainly. On the other hand, they may see it as a risk expense worth amortizing (lowered Bill O' Materiele) as part of their xCloud streaming endeavour if that makes sense, and power efficiency is the name of the game for scaling up servers across the world.
As I've brought up a while back, there would be concerns for die area for phys I/O, so again it'd be interesting in the context of a major overhaul (or not) as to what they end up doing because of the memory issue.
True, crossgen could be around for a while, especially given the enormous success of the PS4. For a 7nm shrink to make any sense you'd need a physically smaller memory bus, which would appear to count out a 256-bit bus. If so, that leaves you with GDDR5 + esram, or GDDR6. GDDR6 would allow you to forgo a clamshell arrangement once 2GB chips were available, but that's a step up in cost over GDDR5 and a big one over DDR3/4.
With Scorpio MS doubled the number of memory channels per GB, which might help with performing both esram and main memory operations from the same memory pool.
Still, I can't help thinking that 16nm is the way that MS will go. Even Nvidia don't have a 7nm product yet, and they can spunk incredible sums on tiny numbers of ultra-margin products.
It's the catastrophic pitfall of the shared bus that's a concern. The game code just runs, so the swaps can happen specifically. I'm not sure how you try to fake that into not transferring or waiting (extra latency really bad for the CPU now) on the memory controller - I guess that is what I'm wondering (since ESRAM is a seamless part of the memory page table) ?
This is a bit over my head (so many things are!), but I can't see why you couldn't intercept the swap / DMA at the hardware level and turn it into a virtual memory address change, and just make sure you wait enough cycles and return the result that the game is expecting. To the game, the page table would look the same as on the base unit, the result would look the same, the timing (in as much as it needed to be) would appear the same. Only difference is nothing has actually moved, and you haven't used any BW.
(Do we know if the DMA units are clocked up in the S along with the GPU and esram? It could be that the game environment is already pretty robust in regard to varying speeds.)
I've no idea is Scorpio does anything like this though. It's just a guess. Scorpio has so much BW and so many memory channels that it can probably burn BW in copies and just laugh it off - I just don't think you'd need to if the system was built with mimicking esram copies and transfers in mind.
Actually, come to think of it, the larger cache on Scorpio probably helps mitigate the pressure of losing the bi-directional esram bus. A larger cache might be something you'd want to keep in a hypothetical "Scorpio Lite".
