AMD Accidentally Releases FSR 4 Source Code, Hinting at int8 Support for Older Radeon GPUs

AMD recently released an updated version of its FidelitySDK, featuring support for FSR 4 upscaling and FSR 3.1.5 frame generation. However, during the release process, the company inadvertently published the full source code for FSR 4 on GitHub. Before AMD removed the repository, several media outlets and social media users, including Videocardz, captured screenshots of the files. The leaked code reveals that AMD was developing a second variant of FSR 4, specifically designed to run on the int8 numerical format. The files show two distinct sets: one optimized for FP8 and another for int8. The existence of a separate int8 version strongly suggests AMD was exploring ways to extend FSR 4’s compatibility to older Radeon GPUs, particularly those based on the RDNA 3 architecture. This is significant because the current version of FSR 4 is not GPU-agnostic—it’s designed exclusively for RDNA 4 GPUs, which feature upgraded AI accelerators capable of efficiently handling FP8 workloads. The int8 version would allow FSR 4 to run on older hardware, potentially expanding its reach to a broader range of users. While the leaked files confirm AMD was actively working on this effort, they don’t reveal how advanced the int8 version is or how well it performs. If it were fully developed and ready for release, AMD would likely have announced it. The fact that it was only partially exposed suggests it may still be in early stages or not yet stable. Adding support for older GPUs would benefit both AMD and the wider PC gaming community. It would allow more users to access the latest upscaling technology without needing to upgrade their hardware. However, the int8 version is expected to deliver lower image quality compared to the FP8 version, due to the reduced precision of int8 calculations. This trade-off is already seen in other upscalers, such as Intel’s XeSS, which offers a high-performance version for Intel’s XMX cores and a more limited, GPU-agnostic variant using DP4a instructions that results in visibly lower visual fidelity. While the int8 version of FSR 4 may not be ready for prime time, its existence indicates AMD is exploring ways to balance performance, compatibility, and quality. Whether this effort will result in a public release remains uncertain, but the leaked code offers a rare glimpse into AMD’s ongoing work to extend the reach of its next-generation upscaling technology.