I'm still trying to reason on how having fast drives is going to shape up this gen.
The big question is if you were to have a drive that had infinite throughput, what would bottleneck your reads first: cpu, gpu, available RAM, bandwidth, optical disc capacity. Assuming the PS5 specs are real, ~9 TFLOPs RDNA, with something like 500 GB/s bandwidth and an 8-core Ryzen CPU around 3.6 GHz (no idea what the leaks are saying), what's the limit?
Starting at the optical disc, everything has to fit on disc. There are time and budget constraints to producing content (models, animations, textures, sounds) and there are limits to how much the data can be compressed. Once it's on your SSD, you have limits in terms of decompressing the content, which could impact the CPU or require a secondary processor to handle decompression. Assuming your cpu or decompression can work fast enough, bandwidth should not be an issue except for bus contention. Both the CPU and GPU are contending for memory, so you can't just hit the SSD non-stop to write into RAM.
Then you have to consider that everything has to fit into RAM. The SSD isn't going to be fast enough to unload everything that's not visible. You can turn a 180 degree view in a frame. So at 30 fps, with an infinitely fast drive you would never have to worry about streaming in an entirely new set of meshes, textures for all of the things within your view. In reality you'll probably be able to read 230 MB/frame max (assuming 7GB/s) drive, but whatever is read won't be decompressed instantaneously, so you probably have some fraction of 230 MB/frame (7 GB/s * 0.033s) multiplied by whatever your compression ratio is (assuming you can decompress within 33ms). I don't know what those numbers would be. There's some limit to how much data you can swap on the fly within a frame. Then you end up back into a coarser streaming scenario, where you're back to streaming in tiles with a caching strategy.
Then you have to understand what the occupancy of your memory looks like. You have 16GB of RAM, minus some system reservation (or maybe not if there's dedicated system memory). Within the game engine there's going to be some percentage of RAM that's lost to the things that we cannot see on screen or hear. It should be on the very small side, but it is a loss. On top of that you need to track things persistently in most games. So the character that walks off screen needs to be there if you turn around. The world simulation does not end outside the view port (in most 3D games). It may be reduced, but not erased. So for every ai, sound, physics object, statistic, the state must remain even if you do not see it on screen. If you increase the density of your world tiles, you're reserving more RAM for the things you can't see yet. Yes there's LODs and whatever, but it's pretty clear that even with an incredibly fast drive, there's going to be some limit to how much data you can stream in per frame, so they'll still have to worry about keeping things resident in memory ahead of its use. There are probably other things to consider in terms of what could actually fit in RAM, but that's what comes to mind off the top of my head.
As for the gpu and cpu, almost everything you'd want to increase in visual density will come with other costs in terms of physics or ai. Unless we get the same static games we always got, you'll have to have some combination of physics and intelligence for every new object you add into the scene. On top of that, you can't just add infinite meshes into your scene before the gpu starts to choke because it hits some kind of vertex processing limit. Something like Nvidia mesh shaders could help massively here. CPU and GPU could both be responsible for physics. Ai is probably still on the cpu, but we've seen a large bump and not all games require huge numbers of ai actors or sophisticated ai.
It'll be really interesting to see what developers come up with, because I think this will be the first time that streaming performance from the drive will not be a huge bottleneck. We went from slow optical discs to 20 MB/s to now potentially many GB/s.
Oh, and I've never mentioned the impact on cpu and gpu caches.
