Disagree, I can find ways to use just the 10G. All texture operations would go there. The OS is going to have some of the lower speed memory. I'd use the rest for cache. It's essentially like the PC's memory structure with separate GPU RAM.
Nope, it's not. If it was the case it would be much better.
Each request that comes for the "slower" pool will make all the memory work as a "slow" pool.
I.e. you will need to ration CPU access pretty heavily.
Obviously "slow" here is not really slow as 336GB/sec is a good speed.
And PS5 will have the same problems if CPU accesses the memory too frequently (as it was the last generation), but the drop in speed will be less pronounced.
BTW, I think that's why MSFT went with 320bit bus, instead of 256bit (like Sony): they have more CUs and they need to feed them, yet RAM bandwidth becomes bottleneck, and on top of that L2 cache in RDNA is sized per each 64bit dual-channel block.
If they could pull off 20GB@320bit that would be nice, but probably costs were so high they needed to compromise.