AMD accidentally leaks FSR 4 source code, teasing possible support for older Radeon GPUs

[Daniel Sims]

Cutting corners: AMD's FSR 4 upscaling is far more competitive with Nvidia's DLSS when compared to previous iterations, but it is not open source and only supports the company's latest graphics cards. A recent GitHub misfire has leaked plans to address those shortcomings and expand the library of FSR 4 compatible games.

AMD accidentally uploaded FSR 4's source code to GitHub this week, leaking files for a version that supports Radeon RX 7000 and earlier GPUs. The company quickly deleted the repository, but not before Videocardz and other observers took a peek.

The leak contained files for an in-development INT8 version of FSR 4, which would have run on AMD's RDNA 2 and RDNA 3 architectures. FSR 4 currently only supports RDNA 4 since it is AMD's first architecture to include hardware based machine learning.

But there's so many copies of it out there that there's little point hiding what's in them, except from maybe the unfinished attempts at an int8 FSR4 model.

Confirmation for sure that AMD was interested in creating a watered down version of the model for less capable hardware.

– uzzi38 (@uzzi38) August 20, 2025

The company's new upscaling method allows it to more closely mirror Nvidia's DLSS, and our testing shows that the two deliver somewhat similar image quality. FSR 4 improves texture quality, reduces blur, and minimizes ghosting in 1440p and 4K, surpassing DLSS 3.

DLSS 4 still remains the strongest upscaling option, but FSR 4's improvements significantly increase the value of AMD's RX 9000 GPUs.

A lighter version of the technology would likely not perform as well, but prior unofficial experiments suggest it could still benefit users with older Radeon GPUs. Using an elaborate Linux based workaround, modder Virtual Cobbler 9930 demonstrated that FSR 4 on aa Radeon 7900 XTX comes with a higher performance cost than FSR 3 but produces a much better image and still outperforms native rendering. It remains unclear how close AMD came to finishing the INT8 version of FSR 4 or whether it will ever be released, but the concept appears sound.

Also check out: AMD FSR 4 vs Nvidia DLSS 4 at 4K

Another significant FSR 4 update that AMD announced this week, which is still on the company's website, is the launch of FidelityFX SDK 2.0. This changes how developers integrate AMD's upscaling technology. FSR 4's machine learning will serve as the foundation for future development.

While users can manually enable FSR 4 in any game that supports FSR 3.1, FidelityFX SDK 2.0 will help developers to automatically transition titles to native FSR 4 support for both upscaling and frame generation in future driver updates. This upgrade will help AMD close the gap with Nvidia's much larger library of DLSS supported games.

Permalink to story:

AMD accidentally leaks FSR 4 source code, teasing possible support for older Radeon GPUs

 

[ferrellsl]

Doubtful that such a release is "accidental".

 

[Liranan]

What better way to create hype than to 'accidentally' release the source code?

 

[Puiu]

They need to accelerate development and game support.

 

[Burty117]

Honestly FSR 4 is actually really good, it brings them right up to Nvidia's standards now.

Something that recently really impressed me, Cyberpunk 2077 on SteamDeck, looks a mess by default, FSR 2.1/3 even in "quality" mode looks god awful.

Intel's XeSS 2.0 in ultra quality mode though? Several magnitudes better, I was actually able to play the game without it being a smeary, blurry mess. Hats off to Intel, you're doing a better job than AMD and you're later to the game.

 

[MasterAce]

RDNA3 does support hardware AI, just not the FP8 format. And it does so in a more hybrid format, where RDNA4 has a ”dead weight” AI core that does nothing unless FSR4 is enabled.

The reason the Linux hack runs slow is that it is running FP8 in FP16, and it takes resources away from other tasks.

 

[Beerfloat]

Amateur hour at AMD as usual

 

[opckieran]

IIRC there’s nothing really stopping AMD from backporting FSR4 to older GPUs, there’s just no performance benefit when running it on them.

 

[Puiu]

IIRC there’s nothing really stopping AMD from backporting FSR4 to older GPUs, there’s just no performance benefit when running it on them.

There is an image quality benefit and a slight boost in FPS. If they manage to improve performance a bit while emulating FP8 compared to what we've seen modders do then that would be really good. (it works on linux)

 

[Theinsanegamer]

RDNA3 does support hardware AI, just not the FP8 format. And it does so in a more hybrid format, where RDNA4 has a ”dead weight” AI core that does nothing unless FSR4 is enabled.

The reason the Linux hack runs slow is that it is running FP8 in FP16, and it takes resources away from other tasks.

I'm not sure that dead weight is the best descriptor. It's a dedicated hardware block just like AVX hardware is.

 

[human7]

Doubtful that such a release is "accidental".

What better way to create hype than to 'accidentally' release the source code?

While unethical, potentially illegal, a competitor could look at the source code to improve their own products, and such a claim would be hard to prove unless they had their own leak. So no, I doubt this was intentional.

 

[Puiu]

While unethical, potentially illegal, a competitor could look at the source code to improve their own products, and such a claim would be hard to prove unless they had their own leak. So no, I doubt this was intentional.

Unless it's something very obvious, these things are usually just mistakes by rookies who upload on public githubs instead of what they should have done. I've seen experienced devs do this this because they forgot to change something when they pushed the code

 

[MasterAce]

I'm not sure that dead weight is the best descriptor. It's a dedicated hardware block just like AVX hardware is.

These days we can’t even run solitaire without DLSS, so ”dead weight” is a bit over dramatic. But it is a die block that does nothing if DLSS is turned off.

Unlike RDNA3 matrix multiplication or AVX, which are integrated into the generic pipeline and share resources like registers with other tasks.

 

[Burty117]

These days we can’t even run solitaire without DLSS, so ”dead weight” is a bit over dramatic. But it is a die block that does nothing if DLSS is turned off.

Unlike RDNA3 matrix multiplication or AVX, which are integrated into the generic pipeline and share resources like registers with other tasks.

Well there's DLSS, DLAA, Frame-Gen (even if I personally don't use it), Ray-Reconstruction, RTX Video Super Resolution and RTX Voice.

It's hard to say it does nothing, unless you exclusively play only games that don't support DLSS/DLAA/Frame-Gen/Ray-Reconstruction, don't use a microphone or watch YouTube.

 

[WaphleStomp]

These fake "accidental" leaks are ridiculously silly. Given the apparent frequency of secrets leaking one would have to assume a kindergarten class would be more secure than whatever it is the big-tech companies are implementing.

 

[OortCloud]

Honestly FSR 4 is actually really good, it brings them right up to Nvidia's standards now.

Something that recently really impressed me, Cyberpunk 2077 on SteamDeck, looks a mess by default, FSR 2.1/3 even in "quality" mode looks god awful.

Intel's XeSS 2.0 in ultra quality mode though? Several magnitudes better, I was actually able to play the game without it being a smeary, blurry mess. Hats off to Intel, you're doing a better job than AMD and you're later to the game.

Now that Intel have their drivers sorted they just need to release some new reasonably priced mid and low range boards and they'd be right there.