Been reading the XSX DF deep dive article this morning and noticed two things I missed previously.
The first thing is that the XSX SoC is what will be used for xCloud servers. I thought they might have used the Lockhart SoC (Sparkman?) first and then upgraded to the XSX SoC later after cost reduction. The second thing is the RAM arrangement. It states 10GB (5 2GB chips) is GPU optimal @560GB/s and 3.5GB/2.5GB is standard running @336GB/s for CPU/OS. What happens to the other 2GB chip? Why not do 5 of each 2GB/1GB chip? Probably missed something here but thought I'd ask.
It's six 2GB chips four 1GB chips. 10 chips total.
In order to get the full 560GB/s you have to be accessing all 10 chips at the same time.
So the 1st 1GB of all 10 chips = the GPU optimal memory.
The slower 6GB of memory is 100% on the 2GB memory chips, but not the entirety of the 2GB chips, just the second 1GB after the first 1GB which belongs to GPU optimal memory.
In other words, when Series X is only accessing the slower pool, it's only accessing the six 2GB chips, but the second 1GB portion of memory as the first 1GB is reserved for the faster GPU optimal memory.
Basically Microsoft reserved specific memory addresses on the 16GB pool of RAM to be strictly for the higher speed 560GB/s and another set of memory addresses (all from the 2GB chips) for the slower standard memory. The reason it is this way is because there is a mixture of 2GB and 1GB chips. If it were all 2GB chips, this wouldn't be an issue, but then Series X would have 20GB of RAM instead and no doubt cost more to produce, meanwhile memory bandwidth would be the same 560GB/s.
I think what you're asking is why don't Microsoft just use five of the 2GB/ chip to get their GPU optimal memory? If Microsoft were to do it that way then memory bandwidth for GPU optimal memory would only be 280GB/s. Each memory chip on the 320 bit memory interface is producing 56GB/s of speed. 5 * 56 would get you only 280GB/s.
So in order to get far more bandwidth, the 1st 1GB of all 2GB chips and the entirety of the 1GB chips is how Microsoft instead gets their 10GB of GPU optimal memory cause then that is 10 chips * 56 for the total 560GB/s.
Basically, one important way Microsoft could use the slower 6GB (or 3.5GB left over for games) is that they can keep data in the slower portion that they plan to eventually use in the faster GPU optimal and just quickly change the memory address of where the data is. Essentially the 3.5GB left over for game becomes a super fast data cache so to speak that's way faster than their SSD (or any SSD).
This is likely made all the easier by the fact that the slower 6GB is on the same exact memory chips as the 10GB GPU optimal memory. Remember, there six 2GB memory chips in total in the system and the slower portion is on the second 1GB section of all six 2GB chips. Basically, it will work out very well because it's a console. On a PC it would be very problematic because developers couldn't more specifically target how they use the memory. On a console it's a non-issue. It's something they have to consider when designing, but it won't actually be an issue.