AMD just accidentally posted the FSR 4 source code, could run on old Radeon GPUs

Someone at AMD is going to be getting a dressing down, as the AMD FSR 4 source code has just been inadvertently posted on Github, and then swiftly deleted. Not only has this slip-up potentially given competitors such as Nvidia the chance to peer at the inner workings of AMD's new upscaling tech, but it's also revealed the presence of files that could enable you to run FSR 4 on older AMD Radeon GPUs based on the RDNA 3 architecture, such as the Radeon RX 7800 XT.

I'm a big fan of AMD FSR 4, which I tested for my AMD Radeon RX 9070 XT review. Unlike previous versions of FSR, it uses AMD's AI cores, and the difference is night and day for game graphics. The previous problems with blurriness and horrible digital noise around fast movement are gone, and the result is much sharper graphics in games, with a big boost in frame rates compared to not using FSR at all. This new tech helped propel several of AMD's latest GPUs onto our best graphics card buying guide, and AMD has also previously told us that there's no technical reason why you couldn't run FSR 4 on Nvidia GPUs using the Tensor cores.

Spotted by X (formerly Twitter) users CaptMcShotty and uzzi38, the FSR 4 source was available for all to see in the FidelityFX software development kit (SDK) section in AMD's GPUOpen Github repository, and there are now many references to deleted files in the commit history where AMD has scrambled to get rid of it all. There are some entertaining comments at the bottom too.

The mistake appears to stem from AMD's announcement of the public release of FSR 4 for game developers on August 20, 2025, as part of its FidelityFX SDK 2.0 suite, via GPUOpen. There's a big difference between making your latest upscaling tech available in a public SDK for game developers with some limited source code and releasing the whole source code for it, however.

One of the areas that CaptMcShotty has highlighted is a collection of folders that have "i8" in their names, as well as references to various FSR 4 quality presets. As CaptMcShotty says, it's likely that these are lower-precision INT8 (8-bit integer) files, which means they could be run by the less powerful AI cores on AMD's last-gen RDNA 3 GPUs.

Just in case you're not up on your data formats. An integer is a whole number, while floating point refers to numbers with decimal points, and the latter is massively more precise when it comes to number crunching. Currently, FSR 4 uses FP8 and FP32, as supported by the more advanced AI cores in its latest GPUs, where FP stands for floating point, but it looks as though there's a way to get FSR 4 running with INT8 data formats too.

I've reached out to AMD for comment on how this happened, and whether FSR 4 could be coming to older Radeon GPUs, and I'll update this story as and when I hear back.

In the meantime, if you're thinking of upgrading your GPU, check out my AMD Radeon RX 9060 XT review to see why this GPU is currently the best budget buy, while our best gaming CPU buying guide will tell you which processor to pair with it.

Have you tried FSR 4 yet, or are you sticking with Nvidia for this generation? You can join our Discord channel to discuss the pros and cons of the latest upscaling tech with fellow PCGamesN readers.