After a two week ban, I'm back. Did I miss anything? :)
(JK, I kept up with the thread)
I think Microsoft already "told" us their memory configuration in the CGi video. They have x10 14gbps certified chips running at an unknown clock. That means 10GB (if they go all 1GB chips) or 20GB (if the go all 2GB chips) or anything in between (if they mix the chip types.
Regarding bandwidth, if the CGI trailer means anything, and the X CGI trailer was very accurate, they have a 320-bit interface and thats written in stone because E3 2019 was already too late in order to change the APU silicon. From here we have 3 options:
1) They are still using 14gbps certified chips. Consoles always down-clock the memory a bit so if we look at how much the X down-clocked its GDDR5 memory, we can expect something around 13.6gbps which will result in around 544GB/s.
2) They will upgrade to 16gbps, after the same down-clock will result in around 622GB/s.
3) They will upgrade to 18gbps, after the same down-clock it will result in around 700GB/s.
IMO 544GB/s sounds enough. 5700XT has 448GB/s bandwidth and even after the overuse of bandwidth by the CPU in a shared memory setup, a Zen2 will never use 100GB/s. So a 544GB/s bus will provide more bandwidth than the 5700XT has, sounds great to me even for a 10TF GPU. But hey, if Microsoft are willing to spend more money in bandwidth all power to them, developers will use it, but it sounds unnecessary.
It's really hard to know what clocks are possible on a new architecture that we knew nothing about. It's as hard to guess what clocks are possible on a new node like 7nm before its been really used. Guessing what clocks a new architecture AND a new node can do is virtually impossible. I guessed that the next gen GPUs will be around 1550Mhz and I thought I was very optimistic, 2000Mhz is a real shock. So I don't think you should blame
Colbert for thinking of lower clocks, for anyone that knows hardware a 2Ghz console sounds like sci-fi.
The demo was 1080p/30fps while RT was used for reflections only. Another thing to note is that the reflections are perfect reflections on smooth surfaces in an urban area without objects with small holes in them like fences and trees which makes it the best scenario possible for RT. So I wouldn't take that demo very seriously, it's very miss leading.
ROPs sit outside the CU, not inside of it. So a 36CU GPU and a 44CU will have the same amount of ROPs. ROPs, just like the CU performance, are affected by clock speed. So even though a 36CU GPU running at 2170Mhz and 44CU GPU running at 1775Mhz have the exact same TF performance, the 36CU GPU has a much better ROPs performance because both have 64 ROPs, but on is running at 1775Mhz and the other at 2170Mhz.
People love to talk about TF, but it's only one parameter in a GPU performance.
