Right, but there's an implicit assumption being made there with respect to the bandwidth requirements necessary for the forced AF or even to deal with the bandwidth contention (note that full 320GB/s is available even for unpatched games even if half the GPU is not), which was what I was getting at. Because the games are unpatched, numerous titles are operating under the assumption that there is a split memory pool, and ESRAM <-> DDR3 swaps may occur on the fully shared bus that may lead to a pitfall in performance if there's not enough headroom. How much overhead they need is not discussed, which is what I was concerned about.
I don't *think* BW would be an issue - internally the esram is very fast for simultaneous read/write operations, but there appear to be very few cases where it allows the X1 to outperform, say, PS4 over the course of a whole frame. And that's even without Delta Colour Compression - Scorpio / PS4Pro can do a lot more with the same bandwidth in terms of ROPerations.
In terms of esram <-> DDR3 transfers, I would assume that said transfers are mimicked by by swapping around a virtual address and pretending a 'real' transfer has taken place. Might even be faster and lower overhead than the real thing!
If the AF overhead was too much, you could always tone it down a little. Witcher 3 (upatched) ran with a 40% bump in res, with higher frame rates and high AF, so Scorpio seems like it would have the overhead to be cut down somewhat and still exceed the OG X1.
Just speculation of course, but I do think there's a lot of potential for a cut down 1X to handle worst case scenarios and run a whole load better 99.9% of the time.
I do think that it would be worth an gpu upclock in performance which should overall mask any deficiency that it would have against esram.
Using ddr4 would probably require more R&D, where they already have stats and actual silicon for gddr5.
One of the problems with a GPU clock bump is that it doesn't get around the 1X's biggest issue, which is that if you're putting buffers in DDR3 (because esram is so pitifully small) you're crap out of bandwidth. Just as a thought experiment, if we were to speculate that out of the 68 GB/s of DDR3 BW, if you actually acheived around 55 GB/s typically in real world, and 30 GB/s of that was generally being lost to the CPU, you'd have about 25 GB/s left for GPU operations.
Now if you were to switch to DDR4 3200, that would be 102 GB/s theoretical, about 80 GB/s typically, and 50 GB/s for the GPU. That would be a huge boost for games bottlenecked by the DDR3. If you were to also add DCC with the ROPs, the thing would fly.
Actually, I wonder if you could use 128-bit GDDR5 with four channels and a clamshell arrangement? It'd certainly give you more BW than DDR3 (88GB/s at PS4 speeds) for decent performance uplift and it would allow for a very small footprint on the PCB.
Would a shrink and keeping ddr3 for the next 3+ years be worth it.
Or no shrink because xcloud is only for infrastructure testing and 1S(ADE), will also be dropped soon after next gen release?
I don't think a shrink to 7nm is worth it if the console is only going to last as long as the X1. With a 256-bit bus you'd struggle to get the area advantage the process offers, and the development outlay would be huge. If you're going to shrink a current system, X1X makes more sense as it's a big chip with a high power consumption. But I'm not sure that a shrunk 1X would be cheap enough to push out as a discless Live / Game Pass machine.
)
