• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

AMD FSR 4 Source Code Leaked - Probably Hackable for Older GPUs

erek

Fully [H]
2FA
Joined
Dec 19, 2005
Messages
17,433
"There have previously been hints that AMD FSR 4 could be coming to older RDNA hardware, and there have even been mod projects that have successfully forced it to work on older hardware, but this is a substantial official signal that AMD may be working on porting it to older hardware. There is also speculation that the INT8 libraries may be destined for the PlayStation 5 Pro, but the file type seems to indicate otherwise—PS5 uses .PSSL, while these shaders were packaged as .HLSL, suggesting they were for PC use."


Official GitHub Link: https://github.com/GPUOpen-Librarie...mmit/01446e6a74888bf349652fcf2cbf5f642d30c2bf

1755813139119.png

Source: https://www.techpowerup.com/340179/amd-fsr-leak-accidentally-tips-potential-support-for-older-gpus
 
They said screen shots were taken but did anyone actually get the source code?
 
Repo got forked and downloaded before AMD deleted it... and came with an MIT license.

Now that the code is loose I wonder if someone will get FSR4 working on NVidia and Intel hardware.
 
must be a sick in your stomack moment when you realized what you just pushed....
 
Repo got forked and downloaded before AMD deleted it... and came with an MIT license.

Now that the code is loose I wonder if someone will get FSR4 working on NVidia and Intel hardware.
I'm hoping that Valve wakes up and makes their own upscaler based on FSR4.
 
I'm hoping that Valve wakes up and makes their own upscaler based on FSR4.
They could do fsr 4 on the new Fremont console as it has rdna 3 at a minimum

Rdna 2 doesn't have the tops nor hardware to support the code pushed accidentally by AMD

AMD was trying to get FSR 4 to work on rdna 3. No chance of rdna 2, imo.
 
Probably not even worth it performance wise.
"the ML-based upscaler of FSR 4 will begin to bear down heavily on an RDNA 3 GPU, which lacks support for many of the high-throughput data sets that RDNA 4 GPUs (have)"
 
Probably not even worth it performance wise.
"the ML-based upscaler of FSR 4 will begin to bear down heavily on an RDNA 3 GPU, which lacks support for many of the high-throughput data sets that RDNA 4 GPUs (have)"
what if someone modded the source to use transformers instead
 
int8 while slower at least started to be something with rdna2-3 and could mean a pure CNN model instead of a transformer-hybrid like fsr 4 probably had at launch
 
  • Like
Reactions: erek
like this
int8 while slower at least started to be something with rdna2-3 and could mean a pure CNN model instead of a transformer-hybrid like fsr 4 probably had at launch
wonder if we'll ever see SSMs or GNNs take the place of transformer-hybrids in this technology in the future
 
They could do fsr 4 on the new Fremont console as it has rdna 3 at a minimum

Rdna 2 doesn't have the tops nor hardware to support the code pushed accidentally by AMD

AMD was trying to get FSR 4 to work on rdna 3. No chance of rdna 2, imo.
I was thinking more of retooling FSR 4 to work on basically any GPU, based on what hardware capabilities it has. Or take FSR 4 and fork it to a completely new upscaler. Right now upscalers are a mess and there isn't a one stop shop solution to this. FSR did work on any GPU, but not FSR 4. If I were Valve and I wanted to make sure my customers aren't confused as to what upscaler they should use, then having a universal one would make things easier. Lets be honest in that if you aren't a [H]ard user of electronics, then you wouldn't know what an upscaler is or which one to use. This would be the best moment where Valve could step in and retool FSR 4 to be used on any GPU and take advantage of newer GPU's based on the hardware they have.
 
The easiest port might actually be getting FSR4 running on RTX 4000 and 5000 cards. From digging around the web it looks like NV added FP8 to their tensor cores starting with Hopper and Ada Lovelace.
 
I remember something, may be HU say they talked with AMD at a conference and the reason FSR4 is not on older cards is because it doesn't work well due to missing acceleration hardware? It's not that they can't technically not get the code working but the cards themself are not setup to properly use it, or something like that, and the experience is poor.

Again, that's just what I believe I remember reading several months ago.
 
The easiest port might actually be getting FSR4 running on RTX 4000 and 5000 cards. From digging around the web it looks like NV added FP8 to their tensor cores starting with Hopper and Ada Lovelace.
While the users using these two generations of cards far exceeds the Radeon user base in its entirety, the instances an RTX 40 or 50 series user would even consider running FSR4 is miniscule and reduced to only game that don't natively support DLSS; not to mention this list of games grows shorter by the day.

So, all in all, nothing of value was lost?
 
reduced to only game that don't natively support DLSS
The much smaller subset of game that do not support DLSS but support FSR 4...

I remember something, may be HU say they talked with AMD at a conference and the reason FSR4 is not on older cards is because it doesn't work well due to missing acceleration hardware?
From my very limited understanding of previous rumors (and I think confirmed by the link) yes, FSR 4 use FP8 operation not present on previous hardware, as an hybrid model, they added traditionnal machine learning for image CNN that work for precise element hepled with transformer for whole scene context.

The leak show an INT8 version fallback could be on the way, maybe a less precise transformer or a pure CNN a la DLSS 3.

At first it feel like it will start to be late to matter much, but the rdna 2/3 install base (ps5-xbox) is quite giant and console now seem to have 12 years long life cycle...
 
Last edited:
The much smaller subset of game that do not support DLSS but support FSR 4...
I searched and found no such game. If fact, there are only a handful of games that support FSR versions from 1.0 to 3.1 but not DLSS, but absolutely none that support FSR 4 but not DLSS. Not surprising, since FSR 4 is only supported on a handful of games to even begin with.
 
I searched and found no such game. If fact, there are only a handful of games that support FSR versions from 1.0 to 3.1 but not DLSS, but absolutely none that support FSR 4 but not DLSS. Not surprising, since FSR 4 is only supported on a handful of games to even begin with.
There will be a small handful temporarily. AMD and NV like to be assholes when sponsoring video games, and launch without support for the enemy upscaler. Then a few weeks or a couple months later they add the other upscaler or two. For example Starfield didn't include DLSS at launch, though it was added later. Indiana Jones and the Great Circle launched without FSR support and of course it was added later. Funny thing is both of those games launched with support for Intel XeSS.
 
I was thinking more of retooling FSR 4 to work on basically any GPU, based on what hardware capabilities it has. Or take FSR 4 and fork it to a completely new upscaler. Right now upscalers are a mess and there isn't a one stop shop solution to this. FSR did work on any GPU, but not FSR 4. If I were Valve and I wanted to make sure my customers aren't confused as to what upscaler they should use, then having a universal one would make things easier. Lets be honest in that if you aren't a [H]ard user of electronics, then you wouldn't know what an upscaler is or which one to use. This would be the best moment where Valve could step in and retool FSR 4 to be used on any GPU and take advantage of newer GPU's based on the hardware they have.
WIth DLSS3/4, FSR4, and Unreal 5's TSR: we are basically at a point where the main upscalers are all just about as good as each other.

Another generational refinement for each, and you probably wont be able to see any difference, beyond extremely zoomed in pixel examinations.

The big hanging issue is that AMD's RDNA 2 and 3 are only able to run TSR. So in any game which is not UE5: you have a bunch of customers left behind the curve. With GPUs which are not very old.
 
I remember something, may be HU say they talked with AMD at a conference and the reason FSR4 is not on older cards is because it doesn't work well due to missing acceleration hardware? It's not that they can't technically not get the code working but the cards themself are not setup to properly use it, or something like that, and the experience is poor.

Again, that's just what I believe I remember reading several months ago.
Yes, we went through that whole song and dance with DLSS on Pascal and older generations. People were adamant that didn't matter or that NVIDIA was just being greedy. You know, the same tired accusations.
 
FSR 4 on older devices is kind of a moot issue.

You can hack FSR4 to work on old things the same way you can hack Ray Tracing and Frame Generation to work on old Nvidia hardware. It’s actively a detriment to performance but you can do it. I remember seeing someone getting FG working on a 1060 and it cut FPS down by more than half, but half of them were generated…

RDNA 1 and 2 just don’t have the matrix multiplication capabilities to make it work well, I wouldn’t be surprised though if it worked well on Vega as it is more capable of Matrix Multiplication than later releases. But how useful would it be on Vega I wonder.
 
Last edited:
Interesting but a one game demo doesn't cut it as far as a performance analysis.
 
To be that costly could be the general rule for that old FSR 4 being used there, run FP8 in FP16 in emulation mode, barely faster than native even on the 7800xt is not that surprising.

That how good those upscaler are getting now, even without a significant performance upgrade people are considering them over native, to clean up TSR or other noise.
 
Last edited:
I was thinking more of retooling FSR 4 to work on basically any GPU, based on what hardware capabilities it has. Or take FSR 4 and fork it to a completely new upscaler. Right now upscalers are a mess and there isn't a one stop shop solution to this. FSR did work on any GPU, but not FSR 4. If I were Valve and I wanted to make sure my customers aren't confused as to what upscaler they should use, then having a universal one would make things easier. Lets be honest in that if you aren't a [H]ard user of electronics, then you wouldn't know what an upscaler is or which one to use. This would be the best moment where Valve could step in and retool FSR 4 to be used on any GPU and take advantage of newer GPU's based on the hardware they have.

AMD has some interesting stuff cooking up for the Redstone project

Apparently it uses ML2CODE framework which converts the AI/ML inferencing logic to compute shader compatible binaries

So there is a performance penalty but should work on any (Radeon) GPU

"Hall began by saying something surprising. FSR Redstone is implemented so that it can run even if the GPU does not have AI acceleration capabilities at runtime. Despite being a machine learning/AI-based technology, it does not require AI acceleration on the GPU side."


"FSR Redstone was developed using AMD ML2CODE (Machine Learning to Code), a research project from ROCm. The core part of the neural rendering technology is converted into optimized Compute Shader code by utilizing ML2CODE."


ML2CODE is included in AMD's GPU computing (GPGPU) platform, ROCm 6.1 and later.

Rather than executing trained AI cores at runtime, it optimizes them as existing Compute Shader code and enables native execution.


https://x.com/CaptMcShotty/status/1966462036956660146
 
AMD has some interesting stuff cooking up for the Redstone project

Apparently it uses ML2CODE framework which converts the AI/ML inferencing logic to compute shader compatible binaries

So there is a performance penalty but should work on any (Radeon) GPU
Looks like the above is starting to work with 7000 series cards, which is great news.

Given the driver overhead issues of Nvidia GPUs the 7900xtx is now a pretty viable uptick from the 9070xt for gaming only uses.

View: https://youtu.be/iOndUPO9_NE?si=BhH_R9Gw7hfeifz_
 
Looks like the above is starting to work with 7000 series cards, which is great news.

Given the driver overhead issues of Nvidia GPUs the 7900xtx is now a pretty viable uptick from the 9070xt for gaming only uses.

View: https://youtu.be/iOndUPO9_NE?si=BhH_R9Gw7hfeifz_

According to the video AMD shifted from Int8 to fp8? I think this proves that FSR4 could adapt what kind of hardware would power it. RDNA2 lacks WMMA which means it won't work on these cards, but I feel like if changes were made it would work.
 
According to the video AMD shifted from Int8 to fp8? I think this proves that FSR4 could adapt what kind of hardware would power it. RDNA2 lacks WMMA which means it won't work on these cards, but I feel like if changes were made it would work.
Other way around FP8 (something that would have ran in FP16 mode on RDNA 3) to int8 (something less precise for this work but has native accelerator on older card I think)

One issue is that even in int8 mode, it seem like almost no performance gain in something like cyberpunk to have it enabled in quality mode (and that it feel like is on the absolute best case scenario a 7900xtx....) as always with this working at all was never the question as it could work on a cpu is it fast enough to still be faster than native is part of actually working... but that a title that can gain to remove noise from native rendering TSR, so we can imagine people preferring 100 fps upscaled to 97 fps native.... that space changed fast enough quality wise that we can see people being interested at it even without a significant performance gain.

From the leak I am not sure they will shift it, maybe more have both dll available at the same time (they had both in it), using the FP8 if hardware acceleration is available.
 
Last edited:
FSR 4 is currently locked to GPUs that use RDNA 4 cores, which the Steam Deck doesn't have. However, that's about to change.

Recently, there was a leak of the FSR SDK, which included FSR 4 files that work on older GPUs that use RDNA 2 and 3 cores, which include the Steam Deck. So, we just had to test it out and see how well it works.


https://steamdeckhq.com/news/we-can...eck-for-better-visuals-with-a-power-tradeoff/

there are a couple of conclusions we can come to here. FSR 4 is a massive improvement over FSR 3, taking away almost all the shimmering and looking incredible. However, the upscaler needs more power in this current state, which can make some games that would be used for it unplayable. There are some I can see taking advantage of it, like Hell is Us and Kingdom Come: Deliverance 2, but there are also ones I would probably say to avoid, like Hogwarts Legacy and Cyberpunk 2077.
 
Last edited:
FSR 4 is currently locked to GPUs that use RDNA 4 cores, which the Steam Deck doesn't have. However, that's about to change.

Recently, there was a leak of the FSR SDK, which included FSR 4 files that work on older GPUs that use RDNA 2 and 3 cores, which include the Steam Deck. So, we just had to test it out and see how well it works.


https://steamdeckhq.com/news/we-can...eck-for-better-visuals-with-a-power-tradeoff/

there are a couple of conclusions we can come to here. FSR 4 is a massive improvement over FSR 3, taking away almost all the shimmering and looking incredible. However, the upscaler needs more power in this current state, which can make some games that would be used for it unplayable. There are some I can see taking advantage of it, like Hell is Us and Kingdom Come: Deliverance 2, but there are also ones I would probably say to avoid, like Hogwarts Legacy and Cyberpunk 2077.
based on teh conclusions of FSR 4, maybe FSR 5 will finally leapfrog DLSS
 
ML2CODE
it doesn't require AI or matrix math cores, just as Nvidia's DLSS technology requires Tensor cores. It can run on conventional GPU shaders. "ML2CODE-based frameworks, such as FSR Redstone, can be seamlessly integrated directly into DirectX or Vulkan graphics pipelines with minimal latency. We believe that the ML2CODE solution is the best way, at least for now, to integrate and deploy 3D graphics and AI technologies." Hall says.


DIRECTX
However, Hall does foresee future AMD graphics architectures that will have dedicated AI cores that are compatible with Microsoft's DirectX Cooperative Vector technology. "Cooperative Vector is a DirectX programming model. It's a framework that provides a way for programmable shader units to perform calculations using other accelerators. In other words, it's a framework that allows shader programs to utilize AI accelerators within the GPU. This is an excellent approach for implementing various AI styles.


"However, currently, significant delays can occur unless the GPU has an architecture specialized for DirectX Cooperative Vector technology. It's true that we're actively working on supporting Cooperative Vector," Hall says.



https://www.pcgamer.com/hardware/gr...one-ai-upscaling-tech-can-run-on-nvidia-gpus/
 
The conclusion I came to was that FSR 4 can work on basically any hardware so long as it's tuned for it. AMD bounced between int8 and fp8 which shows how flexible it is. Remember how it sucked for RDNA3 and now it sucks less? Someone is going to tune FSR4 to work on multiple GPU's and have less performance loss.
 
Back
Top