Some interesting quotes from LeviathanGamer2 I thought were worth sharing.
Why the PS5 is trading blows with the more powerful Series X?
"Well first of we have to address how much more power does the Series X actually have over the PS5 and the answer is not as much as most people think, because they are looking at less specs than they even should be, because there are other hardware components of the GPU that they are not even addressing, the big one that I've not seen anyone discuss is the same L1 cache size, because of how AMD GPU's are laid out, the Series X and PS5 have the same L1 cache sizes, except the PS5 has 8-10 compute units attached the 128 KB L1 cache but the Series X has 12-14 compute units attached to the L1 cache which is also 128 KB cache, so the PS5 cache latency will go further on the PS5 than the Series X...this is a very important factor in performance"
On the future of SSD and I/O architecture on PC:
"On the GPU side, you can do GPU based decompression at the cost of some performance, so that's the direction PC is going to go in to mitigate the high decompression rates on console and you can't just go with a faster SSD speed to fix that problem because there is other bottlenecks that you'll introduce. Like you can't just say PS5 has 9 GB/s of decompression so I'll go with a 9 GB/s SSD, you'll have to go futher than that in order to match that. So decompression is going to be forced to be on the GPU and that is going to come at the cost of performance unless dedicated decompression chips get added to the GPU's which will probably happen at some point but currently no GPU has that support"
On the impact and role of the SSD in next-gen:
"For a hard drive, there is what is called sequential read and random reads. Sequential reads is when you read from a list of (memory) locations whereas random reads is when you jump to locations on a drive. On a hard drive every random read has a 10 millisecond delay which for a 30 FPS game is a third of your frame where your hard drive is idle and your streaming in 0 bytes and this is incredibly damaging to performance. So in current gen, developers package large groups so you hit a trigger point in the game and it loads in a 100 or so models into RAM, you might not see it for the next 2-3 minutes but it has to be in that package and they don't want to tank the performance. On the SSD, there is some impact on the random reads but it's in nanoseconds not milliseconds so it's not as punishing and it only impacts one SSD channel while this occurs, so the rest of the SSD can still be streaming."
"For example, the PS4 under the most intensive load is like 80 I/O operations a second, whereas a SATA SSD and this is a SATA SSD can do 40,000 I/O operations per second under heavy load. So now developers can do random reads without having a performance impact, you can now load in a asset because it's now visible on screen or because it finally showed up, or only want to load in half the model, which is kinda what Unreal Engine 5 is doing is they're basically shooting out rays into the scene and calculating "hey we don't need this model because it's too far away so we'll load in quarter of the mode", so what this allows you to do now is you can massively increase polygon count and you can massively increase the amount of textures through Virtual Texturing or Sampler Feedback...and it's only because you have an SSD, if you took a hard drive and put in these consoles (PS5 and Series X) it would cripple their performance...currently no engine game is built around this, because they were built for PS4 and Xbox One in mind, so they have to completely rewrite the engine and it's done very early in development...so this is going to take a bunch of years to get there'...2022 is realistically when we see that transition".