Tuesday, March 24th 2026
Upgraded PSSR Uses INT8 FSR 4 Implementation That AMD Denied Older RDNA 3 GPUs
Shortly after AMD released FSR 4, claiming that the tech was exclusive to the latest RDNA 4 GPUs, the company seemingly accidentally published the libraries that make up the backbone of the tech, revealing that there may have been a version of FSR 4 planned for RDNA 3 and RDNA 2 GPUs. While this open-sourced oops was later used by modders to bring support to the aforementioned Radeon RX 7000 and 6000 GPUs, a recent Digital Foundry interview with Sony's Mark Cerny suggests that the INT8 version of FSR 4 may have been a compatibility version of the upscaling tech that would later make an appearance as Upgraded PSSR (or PSSR 2.0) on the PlayStation 5 Pro and its RDNA 2 GPU.
According to Cerny, "FSR Redstone and the new PSSR have somewhat different implementations due to the underlying hardware, e.g. FSR Upscaling uses 8-bit floating point and PSSR uses 8-bit integer." He adds that "in practice, the same model is used, but it's trained on different data, e.g. if targeting a 2:1 fixed upscale then the training data used is just for that upscaling ratio - and that different training results in different parameters...not seeing too much difference in results, the various flavors in the updated FSR Upscaling really are rather close to the new PSSR." He also mentions that, on PC, because players are generally so much closer to their monitors than living room gamers are to their TVs, the goals of FSR and PSSR differ slightly.
Source:
Digital Foundry
According to Cerny, "FSR Redstone and the new PSSR have somewhat different implementations due to the underlying hardware, e.g. FSR Upscaling uses 8-bit floating point and PSSR uses 8-bit integer." He adds that "in practice, the same model is used, but it's trained on different data, e.g. if targeting a 2:1 fixed upscale then the training data used is just for that upscaling ratio - and that different training results in different parameters...not seeing too much difference in results, the various flavors in the updated FSR Upscaling really are rather close to the new PSSR." He also mentions that, on PC, because players are generally so much closer to their monitors than living room gamers are to their TVs, the goals of FSR and PSSR differ slightly.
31 Comments on Upgraded PSSR Uses INT8 FSR 4 Implementation That AMD Denied Older RDNA 3 GPUs
Guessing FSR Diamond the Xbox branch will have some sort of exclusive to Xbox feature as well.
This is not cool. You really need to let RDNA3 owners use this technology. The hardware clearly supports it..
Really though they should have just kept quiet and left people speculating about magical custom hardware in the Playstation as usual.
Seems they really don't want customers buying their products.
Making it possible to do FSR4 on PS5 may have been a goal, but there is nothing here to strongly indicate that it was the goal.The only things I can see Sony having a hand in are the training material for and tuning of the ML model's weights, which probably won't be as useful for PC anyway. AMD is very much capable of doing the entire thing on their own. And they have, considering FSR in its entirety, including the base, FP8 FSR4 variant.
I am still a firm believer that int8 FSR4 is a project/product of its own, but I do doubt that its cost are greater than the terrible optics of not shipping it.
About HW ability to run it they do not care.
Edit: grammar.1) Its slower than running it on FP8
2) It provides worse quality output (more artifacts)
3) It would need conversions from FP8 to INT8 and back every time the function is called as the model was made using FP8 .. slowing the process down and adding even more latency
As there is not much things made for INT8 there is very little reason to run INT8 on RDNA4.
Or did you just made a typo and meant RDNA3/RDNA2 ?Knowing Sony I would blame Sony.
BTW .. do not forget to thank Sony for killing all PC alternatives of Horizon and the decision to keep them PS exclusive made afterwards.
The whole FSR4 INT8 debacle has left a sour taste in lots of RDNA2/3 owners and might have hurt RDNA4 sales.
AMD support is just hot garbage…
Probably not, but fun to think about.
"The new PSSR would have to be faster than the original one."
www.digitalfoundry.net/news/2026/03/the-big-pssr-interview-with-mark-cerny
Mark Cerny Confirms Frame Generation "Should Be Seen At Some Point On PlayStation Platforms"
www.digitalfoundry.net/news/2026/03/mark-cerny-confirms-frame-generation-should-be-seen-at-some-point-on-playstation-platforms
still in other topics people are asking if they should buy a amd radeon 9070xt when they own a very nice nvidia 3080 graphic card with the best and latest nvidia software available.
amd is unwilling to open source the full hardware specifictations for their amd graphic cards and their drivers. when amd is unwilling to provide features they should fully open source everything and just make the silicion. that implies no firmware. all stuff has to happen in the open source code for any available operating system. there is for sure such a questionable open source license like oracle and others use which limits the reuse for companies but allows it quite easy for the end user.
edit: @Jtuck9 does not change fact, most features are backported to nvidia 2000,3000 and 4000 cards. that is the core criticism. do not take it out of context, when you know quite well how it is meant by many users in many similar topics. nitpicking in details to the end because you are more in the nvdia marketing bubble and can explain every feature to the end point is nice but does not aid the discussion.
edit 2: i understand this topic that way. hardware technology below radeon 7000 or equal radeon 7000 get those integer math based algorithm and radeon 7000 does not. integer means the algorithm does not use comma in numbers. radeon 9000 can handle floating point numbers, numbers with comma. the sony "paystation (tm)*" shows there is most likely no technical limitation but other factors on the amd side. i speculate to give a benefit to the radeon 9000 cards and make radeon 7000 faster end of life.
*I agree with you fundamentally of course and it's been demonstrated that FSR 4 is of noticeable quality difference. I just wonder about the floor needed for things like upscaling (mentioned in those DF articles) and whether you can even decouple things if they are meant to be used together (Redstone). If upscaling wasn't decoupled from Ray Regeneration would the floor be there for frame generation? How does that translate over to previous architectures?! Wasn't there mention of emulation layers for RDNA 2 /3 cards rather than int8? Will we have emulations layers between RDNA 5 and 4?!*
Anyway, people love giving the middle finger to AMD, as was said above. Today it looks more like a traditional initiation ritual: someone declares their love for Nvidia, gives AMD the middle finger, and gets accepted into the group.AMD could name it FSR 3.5 and add the word “Beta” at the end, explaining that it is just an experimental version offered as is to RDNA 2 and 3, or only RDNA 3 owners, and without any official support. Still, people would complain about AMD anyway. Of course, I do not see the same kind of headlines for Intel, which has been promising to open-source XeSS for the last four years. And there are only excuses when it comes to Nvidia limiting frame generation. People love having double standards, and then they cry that life is unfair and powerful people are treated differently than they are.
Anyway, people with an RTX 3080 asking about moving to an RX 9070 XT, I am guessing they consider FSR 4 good enough as a replacement for DLSS and also want some kind of hardware frame generation, not something like the software-based Lossless Scaling. They probably also understand that Nvidia has been using memory bandwidth and/or VRAM capacity as a kill switch for its models (for the last 20 years), so they are looking at affordable, still-fast-enough 16 GB models. Just a guess. They just do not want to pay 900 dollars or more for an RTX 5070 Ti.
If they can't release it officially due to exclusivity deal with Sony, it's still available for those who really need it. It's a no brainer to download files and place them in relevant folders.They "suck" so much that more than 100 million buyers enjoy their APUs in gaming consoles....
I love when people post braindead nonsense.You were trying to articulate something without coming across as angry teenager?Hot garbage is this comment. There are more features to care about. I don't use upscalers at all, but need more VRAM instead for my games played in 4K. This is the main I haven't bought 80 class card, as those were gimped. Nvidia does have better upscalers, but that doesn't impress me at all.
Paid exclusivity
Tell me, please, how many features were backported from 2000 to older cards (RTX 2060 - TU106 - Jan 7th, 2019 .vs. GTX 1660 - TU116 - Mar 14th, 2019), Turing to Turing?
RTX 2000 cards have CUDA shaders, RT cores, Tensor cores. So it will be shame if basic things would not work. Deap Learning AA, Deap Learning Upscaling, Ray tracing, Ray Reconstruction. And there the compatibility (and backporting as well) ends.
If you want to compare Oranges, tell me when will be backported things like:
- Framegen
- Adaptive Framegen
- Shader Execution Reorder
In non-nVidia world RT does not need RT units (Linux users can run it on Polaris based cards), Framegen does not need specialised units from rtx 4000, Adaptive Framegen does not need "Fake Frame Counter". And Machine Learned (or Deap Learned if you prefer that term) upscaling does not exactly need Tensor cores.
And the last point is good enough for sheeps to start their hate campaign.