Series X has a faster SSD and much faster RAM than the PS5. 2.5x efficiency gains for textures means it reduces the stream size of textures by 2.5x so a 10 GB texture only needs to stream (10/2.5)=4 GB instead of 10. At 4.8 GB/s, (4GB/4.8)=0.833 seconds (1/14)

Without XVA, SFS/BCpack, that 10GB texture would take the SX: (10/4.8)=2.08 seconds. So to load 10 GB in 0.833s u need an SSD with the RAW speed of 12 GB/s this is how Xbox is claiming a multiplier on throughput/memory (2/14)

This works with RAM/memory too so the 10 GB of RAM running at 560 GB/s becomes 560 x 2.5 = 1400 GB/s and the 3 GB at 360 GB/s becomes 900 GB/s. So, what about the PS5? It has no SFS equivalent, there is partially resident textures (PRT) but this is software based (3/14)

Compared to SFS which is hardware housing custom texture filters to reduce LOD pop in and using SF to trigger page reads, it is by in far a superior solution and not even comparable to PRT. (Assuming PRT is being used on PS5 to combat SFS) (4/14)

DF said that there was no visible LOD pop in the Hell-blade 2 trailer, thus they concluded it wasn’t real-time, but the above info was not available then. That was their goal with SFS to eliminate pop in and increase perf (Also, HB2 was real time confirmed by UE engineer) (5/14)

Anyway, SX SSD has sustained minimum speeds, whilst there is no indication that the PS5 SSD runs at 5 GB/s 100% of the time, it is only mentioned as a target. So, let’s bring it back to the original calc, loading in 10 GB but with the PS5 tech now. (6/14)

To do this, u must look at Oodle codec, which has a maximum lossy decompression rate of 50% (BCpack reportedly has 50%+). So now the 10 GB becomes 5 GB, thus at 5.5 GB/s: (5/5.5) = 0.909s, slower than the SX which did the same texture in 0.833s (7/14)

With that speed for PS5, to load 10 GB in 0.909s: (10/0.909)=11 GB/s, compared to SX’s 12 GB/s. Note: SFS/BCpack 2.5x increase is only for textures, however textures make up %70+ of a game’s file size. So, for the other 30% of data, PS5 is faster BUT only in some cases (8/14)

Part of the 30% data are polygon models (triangles) and SX uses mesh shading, a hardware-based mesh compression tool. PS5 doesn’t have Mesh shading, it has primitive shading but that was introduced in RDNA 1 cards years ago. (9/14)

Mesh shading was already drawing 50 million triangles on rtx cards at 4k 60 compared to PS5’s 1404p 30 FPS with only 20 million drawn triangles. Mesh file sizes are smaller anyway. “But Cerny said 22 GB/s”, that is a theoretical maximum when decompressing synthetic data (10/14)

Like a string of even numbers, types of data which is useless for games (Kraken is general purpose after all). If we want to talk theoretical maximums, SF alone delivers over 10x improvements if the “data crunches right” in ideal scenarios: 524k vs 52k (11/14)

So, the SX can do 4.8x10=48 GB/s peak, its BS just like the 22 GB/s as they are unrealistic values. Also, I forgot to mention, BCpack is hardware based whilst Oodle is software based. And SFS 2.5x is an average, meaning it is sustained, not peak perf increase. (12/14)

So, while PS5 is faster with 9 GB/s vs 4.8 GB/s for things other than textures/models, their size is still insignificant to textures and were never the bottleneck to begin with. So SX wins in the most important category for SSD’s. (13/14)

This analysis doesn’t even touch the 100 GB of instant access and using storage as VRAM, that’s something else the PS5 cannot do either however the specifics of that are still murky and no one really knows what it means (14/14)

Also huge thanks to @Kirby0Louise for the education, pretty sure I got most of it right lmao

The first part is wrong, SFS is not a multiplier of memory bandwidth, only capacity, my bad (wish we could edit tweets smh) (thanks to @Kirby0Louise for pointing that out) but the 2nd part about PRT remains true.

There IS a silver lining in both of these, if Oodle is implemented on kraken hardware, then the 9 GB/s metric can be used to allow the PS5 to do 13.5 GB/s for textures. If BCpack's compression is >50%, that lead will fall again. All in all I think the differences will be small af