There is no hard partitioning it defeats the whole purpose of unified memory.
The difference is CC (cache coherent) or no-CC access.
CC is needed for more complicated scenarios, where you need to do computations with CPU and GPU on the same blocks. But it doesn't mean that once block is accessed through CC bus it cannot be accessed through non-CC one.
The main benefit of unified pool is flexibility when it comes to assigning system ram and vram as opposed to hard limits
I admit i don't know if whats partitioned as system ram can be accessed by the GPU through the fast bus, if you are certain I'll take your word for it
Anyways i think 8GB VRAM cards will be fine for the first two years of crossgen ports, for next gen only games 16GB VRAM GPUs will be set for the whole gen.
I've done some calculations assuming next gen consoles have 20GB available for games and a 4GB/s SSD:
8GBVRAM/32GB DDR4 PCs would need a 2GB/s SSD
So you were right a 500MB/s SSD wouldn't be enough even accounting for the extra memory buffer, but i don't know if such scenario where you need constant non stop 4GB/s+ streaming will be common
Im inclined to believe a 20GB buffer would be enough for current gameplay till the next loading screen. We'll see i guess, especially interested to see how Sony exclusives exploit it.
It would be a shame if 3rd party games are (ironically) held back by PCs