Regarding memory, I'm just not sure if
32GB and a 256bit bus is good enough, when looking at memory bandwidths. Here are the numbers for memory chips with different speeds:
14Gb/s chips: 448GB/s
16Gb/s chips: 512GB/s
18Gb/s chips: 576GB/s
But what is actually required? If we look at AMD Vega 64, it's memory has a bandwidth of 483.8 GB/s. From what I have read, Vega performance is heavily tied to memory performance, so you probably can't go lower. And Zen (1) seems to
peak at ~50 GB/s. Which probably means you want to have at least
484 + 50 = 534 GB/s memory bandwidth.
Looking at PS4, from what I remember,
it's effective memory bandwidth was actually lower than the peak value, especially when you have CPU and GPU accessing memory at the same time. Which could mean that you want to have even higher bandwidth numbers than just the sum of each part's peak values.
So if we look at the numbers above, it seems 14GB/s and 16GB/s chips won't be fast enough, leaving only 18GB/s chips. And not sure if that leaves enough room for when you have to deal with inefficiencies and parallel access (or if the PS5 GPU is more powerful than Vega 64).
With
24GB and a 384bit bus, those numbers look of course much better, even with "slow" 14 GB/s memory chips, leaving plenty of room:
14Gb/s chips: 672GB/s
16Gb/s chips: 768GB/s
18Gb/s chips: 864GB/s
In the end I guess it's probably a matter of which costs less in the long run: Using a 384bit bus with lower specced memory chips vs a 256bit bus with higher specced memory chips. And also, if those massive bandwidth improvements of a 384bit bus are more useful to developers than 8GB additional RAM.
