Ms said before that the work they make for fps boost and auto hdr can be applied by the developer as its own patch (probably just wrapping to get the X|S seal), I wouldn’t doubt it was they had to do on that version (otherwise they would have increased the resolution for SS to 1080p).
On Ps5 however, it seems to be a proper patch since they modified the resolution in order to achieve the 2x fps boost
Sorry, I should have been specific and mentioned third-party studios. Microsoft is amazing at supporting its games, and it’s a huge reason I decided to go with an Xbox this gen. Feel like I’m getting VIP treatment and free perks out the wazoo
Very interesting insight, thank you for sharing!
This really seems like the pattern and I couldn’t be more pleased. Would be cool if there were another option for people who don’t have VRR displays, but I assume that just isn’t as easy as it sounds. I’m always rooting for more options, but it’s kind of dumb to do so without having any idea of the amount of work it requires.
Obviously, it could be made to run on GPUs without the instructions but at slower rates. The question is if the slower rates are faster than running at the higher resolution to begin with.
I still have questions around the performance timings of XeSS on Intel’s ARC GPUs. I want to see more of that first before moving on to how it performs on RDNA2 GPUs.
this also shows that people who said ps5 is custom RDNA 2 and kept only the best features is full of crap, dot-product instructions are kinda important
They do. They’re matrix engines. However there’s been a lot of chatter in the tech circles how their algorithm may not be as efficient as DLSS. Their slide deck material doesn’t get into specifics or compare to DLSS. So I’m being cautious on how widespread its usability may be. There’s a lot that’s left to be seen how it really plays out.
This is a slide about it, but note their disclaimer about conceptual illustration purposes only.
It depends on speed of execution. You get that by having either Matrix Engines or DP4A instructions. The DP4A instructions are like rapid-packed-math, where they execute faster. If you’re forced to use INT16 operations instead of INT8 you’re going to run at half the speed. It may be even slower if you don’t have the Dot Product instructions.
DP4A computes the equivalent of a total of eight integer operations.
DP4A performs the vector dot product between two 4-element vectors A and B (each comprising 4 single-byte values stored in a 32-bit word), storing the result in a 32-bit integer, and adding it to a third argument C, also a 32-bit integer.
SX and SS supports the version that doesn’t require tensor cores, and while not as fast, still applicable.
Ps5 can run the same algorithm, but without the hardware supoprt X|S has each operation needs 4 cycles on it versus just 1 on the xbox consoles. That may prevent the algorithm from being fast enough on the console to be a viable choice.