Despite ostensibly launching its new RDNA 4 graphics cards and FSR 4 AI-based upscaler at CES 2025, there was much more information released to journalists in pre-briefings than in AMD’s 45-minute CES 2025 keynote. Thankfully, Digital Foundry’s Alex and Oliver got to see both in an AMD suite on the show floor, with a machine learning upscaling “research project” on an RX 9070-series GPU that is almost certainly FSR 4. The game of choice was Ratchet and Clank: Rift Apart, one of our go-to titles for comparing upscaling technologies thanks to its support for nearly all of them. That gave us a great opportunity to see how FSR 4 is shaping up ahead of its launch with the new GPUs later this quarter – and we’ve got to say, we’re impressed.
The advantage here is that FSR is moving from being an hand-tuned analytic upscaler to one that takes advantage of ML or machine learning (one of the many technologies under the ‘AI’ umbrella). We saw a similar evolution from Nvidia between the original DLSS upscaler and DLSS 2, and that brought about a huge improvement to image quality – so is it the same for FSR 4?
In a word, yes. In side-by-side comparisons between two PCs running Rift Apart in 4K performance mode, one with FSR 3.1 and one with FSR 4, the new ML-based technique looks to solve many of the previous versions’ biggest issues.
Most notably, image quality seemed noticeably improved on the “research project” PC across a variety of scenes. The fine texture of the red carpet in the game’s opening level is a hard one to reproduce, for example, with FSR 3.1 compressing a fair amount of detail and producing a moiré pattern, but FSR 4 managing to better preserve the individual fibres.