FSR 4.1 AI upscaling finally coming to older Radeon GPU users

When AMD launched the FSR4 AI-based upscaling technology for games last year, it was exclusively available for the new Radeon RX 9000 generation GPUs using the RDNA 4 architecture. This was despite the fact that a version using INT8 compute compatible with older GPUs had leaked out, apparently by accident. But owners of older Radeon cards are finally in luck now: FSR 4.1 is coming to RDNA 2 and RDNA 3 as well.

AMD has kelong kept silent about the possibility of releasing FSR4—or perhaps some alternative compatible “FSR4 Lite” version—for older GPUs, though it never explicitly ruled it out. VideoCardz, for example, reports that AMD representatives were repeatedly asked via email about plans for FSR4 on older GPUs, but never giving an answer. It now appears that this may simply have been because work was ongoing, but for some reason AMD was not ready to confirm it.

That has now changed. Last year, AMD officially announced that the recently released FSR 4.1 update to the technology will eventually expand AI upscaling support to older graphics cards as well. The first to receive FSR 4.1 will be Radeon GPUs based on the RDNA 3 architecture from 2022 (the RX 7000 generation), which will gain support in July. While not explicitly stated, RDNA 3.5 integrated graphics are presumably included in this as well.

This is expected to function within the same infrastructure already used by FSR 3.1 and current FSR4/4.1 on Radeon RX 9000 cards, meaning a driver update should be sufficient and games will not require individual manual patches to add RX 7000 support. AMD says FSR 4.1 is expected to work “out of the box” in approximately 300 games when the support launches in July.

Podpora AMD FSR 4.1 na grafikách Radeon RX 7000 a RX 6000
Support for AMD FSR 4.1 on Radeon RX 7000 and RX 6000 graphics cards

Although it was mostly speculated that only Radeon RX 7000 cards would likely receive AI upscaling support due to their stronger AI acceleration hardware, AMD is going further. The company has also officially confirmed that FSR 4.1 support will later be delivered to the previous-generation RDNA 2 architecture as well—that is, Radeon RX 6000 cards. For these GPUs, support is scheduled to arrive via drivers at the beginning of next year, likely because additional tuning and testing are still required (the previously leaked FSR4 version did reportedly perform noticeably worse on RX 6000 cards than on RX 7000 hardware).

Podpora AMD FSR 4.1 na grafikách Radeon RX 7000 a RX 6000
Support for AMD FSR 4.1 on Radeon RX 7000 and RX 6000 graphics cards

AI upscaling only?

This is not a leak or rumor, but an official announcement directly from AMD senior vice president Jack Huynh, who serves as the company’s general manager for the Computing and Graphics segment. It is important to emphasize, however, that only FSR 4.1 AI upscaling itself has been mentioned so far in this announcement—not other components. Radeon RX 7000 and RX 6000 cards may thereforre very well not receive the improved AI-based frame generation AMD has since released for Radeon RX 9000 GPUs.

Model will run in INT8 mode

AMD has confirmed that FSR 4.1 on Radeon RX 7000 and RX 6000 generation graphics cards will use an AI model operating with the INT8 integer data type, just like the leaked FSR4 version and Sony’s PSSR2 upscaling used in the PlayStation 5 Pro. Only RDNA 4 supports FP8 (a floating-point data type) calculations, which regular FSR4 technology uses on Radeon RX 9000 cards.

This is not a trivial difference—INT8 has significantly worse dynamic range, which directly impacts AI model output quality. This was already observable in the leaked FSR4 version, where RX 7000 cards produced inferior image quality compared to full-fledged FSR4 because of INT8 limitations.

FSR 4.1 na starších grafikách bude používat výpočty INT8 na první generaci AI akcelerátorů (RDNA 3, Radeon RX 7000) nebo na obecných výpočetních jednotkách (RDNA 2, Radeon RX 6000)
FSR 4.1 on older graphics cards will use INT8 calculations either on first-generation AI accelerators (RDNA 3, Radeon RX 7000) or on general-purpose compute units (RDNA 2, Radeon RX 6000)

According to Jack Huynh, AMD worked extensively on refining and improving the integer-based FSR 4.1 model. This reportedly included memory utilization optimizations and better motion vector selection (which can help prevent errors, artifacts, and motion blur). Proper functionality was also tested across a large number of specific games.

As a result, the final FSR 4.1 release should hopefully deliver better image quality. It is possible that this tuning process is precisely why AMD took so long to officially announce support for older graphics cards—with versions prior to FSR 4.1, the company may not have been confident the public would accept the INT8 version. At the same time, AMD likely prioritized support for new RDNA 4 cards, initially. Whether that was genuinely the case, or whether it was the case of AMD deliberately delaying support to help drive demand for newer GPUs, is something we will probably never know for certain.

Sources: VideoCardz, AMD

English translation and edit by Jozef Dudáš


Contents

AMD confirms development of little “LP” core for Zen 6

When AMD came back nine years ago with the Zen core, its strategy was to eliminate the split between big and little (Bobcat and Jaguar) cores and cover the entire market with a single, balanced architecture. Intel, on the other hand, moved to hybrid (big.LITTLE) CPUs. AMD is apparently going there too, now. It has used classic and compact cores where both shared the same architecture so far, but now there will be a separate low-power core. Read more “AMD confirms development of little “LP” core for Zen 6″ »

ZLUDA project enables PhysX effects on Radeon graphics cards

PhysX effects simulating realistic physics are a good example of the risks associated with proprietary technologies in games. Nvidia kept them as an exclusive feature available only on GeForce graphics cards for years, only to ditch support on the new RTX 5000 series, causing the effects to stop working. Ironically, this may have been the catalyst for an independent emulation effort that, years later, brings them to Radeon GPUs. Read more “ZLUDA project enables PhysX effects on Radeon graphics cards” »

AMD Re-Enables RAM Encryption on Ryzen CPUs Following Criticism

Recently, reports spread that AMD had disabled TSME, one of the Ryzen CPU security features without warning. It’s a bit more complicated than that, however, as this was not a regular default feature, but rather a Pro SKUs technology that was only present on standard Ryzen CPUs unofficially if not accidentally. The media attention seems to have resulted in access to TSME technology being restored for the non-Pro Ryzen processors, though. Read more “AMD Re-Enables RAM Encryption on Ryzen CPUs Following Criticism” »

Leave a Reply

Your email address will not be published. Required fields are marked *