Last month, following the accidental source code leak of AMD FSR 4 INT8 files, modders managed to get a version of AMD FSR4 working on non-RDNA 4 GPUs (the Radeon RX 9000 Series that launched earlier this year). The regular version of the latest AMD upscaler runs on FP8, that is, the 8-bit Floating-Point format, which is not supported by prior RDNA GPUs. However, this leaked version running on INT8 is compatible with a much wider range of hardware, since support for the 8-bit Integer Fixed-Point format is included in most modern GPUs.

Of course, there's a reason why AMD chose the FP8 format for the public release: it allows for a wider dynamic range than INT8, doing a better job in preserving model accuracy and stability. Digital Foundry's Richard Leadbetter has now tested this version

See Full Page