AI oversampling: AMD wants to run FSR 4 on Radeon RX 7000 86 comments
Image: AMD
AMD briefly talked about RDNA 4, the Radeon RX 9070 XT, and the RX 9070 at CES, but a full announcement has yet to follow. In the meantime, more and more information about the new AI FSR 4 supersampling is being made public – but it provides little clarity.
FSR 4 – where is it currently running?
During its own presentation at CES, AMD only mentioned in passing the new RDNA 4 and FSR 4 GPU architecture. The information shared with the press before the show on AMD’s new AI supersampling, expected more later this year, were a little more detailed; However, as with the new Radeon RX 9070 XT and RX 9070 gaming graphics cards, a full presentation is still pending.
However, at the time of the CES keynote, AMD had already given two details on the graphics cards on which the FSR 4 would actually work: on the one hand, the slides shared beforehand with the press explicitly mentioned the “Radeon RX 9070 Series”; On the other hand, journalists were told that the new AI units of RDNA 4 are basic and therefore all new RDNA 4 graphics cards are supported.
FSR 4 is AMD’s first AI supersampling, but only works on graphics cards with RDNA 4 (Image: AMD)
This quickly caused confusion, which AMD attempted to clear up in early discussions at CES following the presentation. Therefore, the second claim, that FSR 4 would be available on all new Radeon RX 9000 series graphics cards, is correct. And the RX 7000 is left out? Apparently, it’s not that simple, as Frank Azor, AMD’s chief marketing officer, explains in an interview. The Spanish formulation was written by El Chapuzas Informatico.
FSR 3 will continue to grow as it works on every GPU on the market as it does not require machine learning. But there are limits to visual quality.
We can significantly improve quality using machine learning, but to do so we need a lot of computing power. Therefore, currently only RDNA 4 cards can run FSR 4. It is possible that we can optimize FSR 4 for the RDNA-3 architecture. And that’s what we’re doing, we want to do it, but we still have a lot to do.
Frank Azor, automatically translated via El Chapuzas Informatico
On the one hand, Azor reiterates that FSR 4 does indeed require high AI computing power, which only RDNA 4 can currently provide in this form. What is new, however, is the time restriction: it is “possible” that FSR 4 will also work on older RDNA 3 graphics cards; but it still requires a lot of work. AMD has already installed AI units with RDNA 3 and the corresponding Radeon RX 7000 for the first time, but this step was only half completed. Until now, the units in question with FSR were inactive, but are now significantly more powerful in the second generation with RDNA 4. AMD is currently working on making FSR 4 work on RDNA 3; The will is there – however, as with FSR 4 in general, Azor does not yet specify a period.
Additionally, it is not clear whether all RDNA 3 graphics cards are actually suitable. It is also said that only selected high-end models have enough reserves to accelerate the neural network behind FSR 4 without much loss in performance and latency. But this is currently still open. Regardless, as a precaution, Azor’s announcement should be understood as a statement of intent, but not as a guarantee. Given the vague information available at the moment, gamers shouldn’t necessarily assume that FSR 4 will be rolling out widely to older Radeon graphics cards – and if they are, they should be pleasantly surprised.
Ratchet & Clank demo gives us a lot of hope
Also somewhat surprisingly, AMD has already presented the new FSR 4 at CES in the worst case oversampling in Ratchet & Clank: Rift Apart (test). Hardware Unboxed and Digital Foundry have now released preview videos and unanimously confirm that AI supersampling has taken a huge leap forward in image quality compared to FSR 3.1, which extends on a whole series of previous construction sites. Both videos are worth watching.
Themes: AMD CES 2025 FSR Gaming Graphics Radeon RDNA 4 RX 9070 RX 9070 XT Those: AMD, Frank Azor, Michael Quesada and El Chapuzas Informatico
An engineer by training, Alexandre shares his knowledge on GPU performance for gaming and creation.
