I'm clearly not a GPU expert so I can just agree for the RT part. I know that's in reality, XsX can't reach most the time his raw advantage and we will obtain similar results on PS5. I think that's the same thing we have for the compute part. For GPU cache part, I totally agree with you, I think that the PS5 has a good advantage at this level.
For the memory aspect, I disagree about the XsX split memory. I'll try to fastly explain why (I said that when I will have more time, I will share past experiences I had with many memory configuration, splitted one between them). In reality, you need to not watch the bandwidth has separated, and consider that you have 560 GB/s "max" bandwidth for the whole memory subsystem. The thing is simply that you have in your pool (for the six 2GB chip) half adresses only dedicated to fast pool (+ the four 1 GB chip, to be used by the most demanding component, the GPU for example) and the other half for the slow pool (for the other components that request less bandwidth and "smaller" datas). It's totally conceptual.
We can take the example where the fast pool is dedicated to the GPU and slow pool for the other components. When the CPU access to the memory banks, it will allocate/use the channels exactly in the same way that is the case on PS5. The difference is that on PS5, 8 channels can burst the data whereas on Series X, it will be limited to 6. On Series X, when the GPU needs to access a "big" data, it can use all the channels to burst data from the fast pool in one transaction (at 560GB/s), and then if the CPU for example need to read a data, it can use max 6 channels to burst data from the slow pool immediately afterwards (at 336GB/s). On PS5 these two steps are done with the same 448GB/s bandwidth. In the two cases, the more the CPU access to the ram, the more it reduces the bandwidth allocated for the GPU, but it will also reduce the OVERALL bandwidth on Series X. That's the only difference between the PS5 and the XsX (in addition to the dedicated pools 10GB + 6BG). But we know that the GPU is clearly the only part that can really use/saturate the memory bandwidth. So the overall bandwidth on Series X won't be too much reduced.
I can add that two flaws are directly possible with this memory configuration (non uniform). A software/game can be bottlenecked by the 6 GB slow pool, in this case that's simply that the ressources are not distributed as it need to be between the two pools. That's the main problem we can encounter with this configuration, where the devs will have extra-work to mitigate that and where we need good tools (yes I know, a forbidden word
). The second problem could simply appear when the datas quantity that require high bandwidth (>> 336GB/s) will exceed 10GB. In this case, I think that "technology" as sampler feedback streaming could help to reduce the quantity of datas stored in the memory (if I'm not wrong).
For me, as I said the splitted memory is not a really big problem. But I can be wrong and we will see.