Insomniac themselves already claimed a ~15% increase in load times on slower SSDs. I'm not sure what the contention is because these numbers are almost completely in line with that.
It performs exactly the same as the PS5 internal SSD in ratchet and clank during gameplay scenarios. That's a pretty incredible 15% jump from 3.2GB/s all the way up to the required 5.5GB/s spec. They will become increasingly more creative with how they take advantage of the PS5 SSD, but something I've always said that I think has been largely overlooked from the very onset about these consoles and their very fast SSD powered I/O was the amount of physical RAM. Assuming the PS5 has a similar 2.5GB OS Reserve as Series X, that's 13.5GB of RAM for games. But let's assume the OS reserve is smaller at 2GB, then that's 14GB of RAM for games. Seeing what that 3.2GB/s drive achieved, the question needs to be asked : Just how much better during actual gameplay will the faster internal drive of the PS5 be able to load data into 13.5GB-14GB of RAM? Every drive was loading those ratchet saves The difference was literal tenths of a second.
Maybe if the consoles had 20-24GB+ of RAM perhaps then I could see a scenario where the faster PS5 SSD has more space in RAM for that advantage to actually show up in a more obvious fashion. As it currently stands, does anyone think Insomniac somehow did not use all the system RAM available to them in Ratchet and Clank?
You know what I like about that, it is that Velocity Architecture = SSD (and decompressors like BCPACK) + Direct Storage + SFS so when people pull out all three they are double counting by definition

.
Not true actually, since Sampler Feedback Streaming isn't automatically in use simply because a dev uses direct storage in concert with the decompression hardware. A game has to be designed and built to use Sampler Feedback Streaming. It isn't as simple as "you're now utilizing direct storage, which means you're also using Sampler Feedback Streaming."
I should know a little something about double counting. I was the person who famously double counted the shit out of some figures in another thread on the subject.
Remember the Dirt 5 technical director said the Series X SSD and hardware can achieve 10GB in 2 seconds without even using decompression?
10GB with Sampler Feedback Streaming's 2.5x efficiency drops that memory demand down to 4GB.
4GB divided by 2.4GB/s with no compression = 1.6 seconds.
However, in the SFS demo if people remember, Microsoft confirmed that the absolute guaranteed minimum speed for the Series X SSD when maintenance related stuff is taking place is 2GB/s.
4GB / 2GB/s = exactly 2 seconds. That is an
effective 10GB of on screen texture data loaded into RAM in just 2 seconds tops solely by using Sampler Feedback Streaming without ever touching compression hardware. The only thing I applied to the calculation was Sampler Feedback's reported 2.5x memory efficiency. My assumption is this is how he achieved 10GB in 2 seconds. At least it matches perfectly.
Series X would still be left with 9.5GB of RAM in this hypothetical scenario. 6GB at 560GB/s 3.5GB at 336GB/s.
Chances are the way he achieved that in his testing was by using SFS without compression. If Sampler Feedback wasn't at all in play in his tests, that's just insane. That guy, by the way, is now at Playground Games working on Fable! So excited to see what they achieve.