Yup. Doesn’t really matter as MS has their own version any the algorithms will be basically identical anyhow. There are only so many parameters to train on.
Found this online too:
never mind, turns out any gpu can use it and does not require tensor cores, wonder how it compares to directML super res. What is XMX and DP4a?
1 Like
XMX is Intels Matrix Extensions for fast computing of ML inference tasks. Its basically their tensor core design but seems more directly integrated in the GPU cores.
DP4a is I believe INT8 packed math
2 Likes
The way it is developed… It will work on amd and Nvidia hardware as well.
Got to know this from hardware unboxed video.
1 Like
So in theory, rdna 2 gpus can use the Intel DP4a to upscale from 1080p to 4k using INT8 operations? Will the final IQ be similar to XMX?
We will simply have to wait and see… it works… But how well? Only results could tell which no one has now.
1 Like
Yup, pretty sure this is right. For computing INT8 dot products real fast. It seems to be new to RDNA2 spec, so I wonder if this is related the logic MS pushed AMD to add to XSS/XSX to make inference task acceleration possible (Gossen mentioned it in his DF interview and it got noted at HotChips too).
I would guess that XSX ML SR might essentially line up with the XeSS-DP4a bar on the chart I posted.
2 Likes
They said it’s open to all gpus (which honestly is what they need to do to stand a chance against dlss)
So tensor cores should provide around 2x more performance for upscaling than RPM for int8, but both still insignificant compared to rendering natively.
That’s great
The algorithm is the same XMX just speeds up it further. (It should be the same for DLSS if it was available on other gpus as well)
1 Like
That is what it looks like in this setup at least. 
1 Like
In theory, yes. All these ML super resolution techniques depend on how fast the gpu computes. If native 4k is faster than 1080p + ML upscaling then nothing is won. It works for NVidias tensor core design because they offer a lot of ML compute. We have to see how practical Intels algorithm works on RNDA2 gpus.
IQ or ML results does not depend which part of the GPU computes the dot or matrix products.
I hope they open source it but i have my doubts.
1 Like
In the presentation they said they optimized the algorithm so it runs great on just int8 dots and that even some of the integrated gpus from them or the competition already supports some acceleration for that.
Tensor cores with full matrix multiplication support will be faster, but they said in either case upscaling should take a minimal time compared to rendering in native res (as it needs to be otherwise there’s no point). Though without RPM for int8 (as in, taking a 4x performance hit) could prevent this from being used on some GPUs, like for example PS5.
They actually promised to open source. They will first integrate as an SDK (and the XMX version is already being sent to engine providers, with the DP4a coming shortly), but eventually will work towards opening it completely.
I think they know that if they want to have even considered over DLSS it needs to at least be way more compatible and open than DLSS and provide at least comparable results.
2 Likes
I love competition.
3 Likes
When it is on more equal footing like now, yeah. 
1 Like
XeSS is open source… Intel will literally pay devs to use it. and if it’s effective MS will definitely use it lol.
1 Like
Sounds like that addition that MS had AMD add to their chip for ML might help accelerate it too.
4 Likes
Another example of DF’s disgusting pro-Sony agenda. Early on, the narrator says PS4 Pro but merely says One X instead of Xbox One X. I hope they go out of business 
All joking aside, is it weird at all that the devs made a patch to a game that’s three years old? I don’t really have a frame of reference for this, but from the Series X side of things, it was pretty much wasted labor since it appears to be at parity with the FPS Boost results. I haven’t played this game yet, does anybody know about loading times? Maybe those have improved too? What about running from an external HDD, is that still possible with this version?
1 Like
hellblade is 4 years old but that got a next gen update to be fair same thing with MCC that is nearly 7 years old and that got updated a week after next gen launched
So, Ps5 increases the resolution output but actually lowers the internal resolution from Pro due checkerboard (unless they are wrong and it’s CB on pro as well which seems likely).
SX keeps the same near 4k image as 1X but doubles the framerate (it seems it just applied the fps boost Ms provided as a patch), so it offers a significant resolution advantage but hammered by lower framerates on the stress areas.
Another case of VRR with SX being the best option where you get the best of both worlds, the smooth presentation without hiccups and better IQ
1 Like