Earlier this week, AMD launched FidelityFX Super Resolution (FSR). FidelityFX is nominally AMD’s response to DLSS, and while it works on different principles than Nvidia’s upscaling network, both features are positioned as performance enhancers that can reduce the amount of work your GPU is doing while increasing overall image quality.
FSR is a single-frame spatial upscaler that analyzes a frame for edges and then resolves them in a manner intended to improve image quality over baseline while minimizing jaggies and shimmer. DLSS is an upscaling model that runs on an Nvidia GPU’s tensor cores. Unlike FidelityFX, DLSS does take temporal information into account and trains on 16K images. FSR doesn’t replace whatever AA solution already exists inside of a game, but the fact that it relies on spatial data limits the type of improvements it can offer.
It’s not possible to compare FSR and DLSS in any game right now, because there is no title that supports both standards. Right now, there are seven games available with FSR: 22nd Century Racing Series, Anno 1800, Evel Genius 2, Godfall, Kingshunt, Riftbreaker, and Terminator: Resistance. Twelve additional games are expected to offer support for FSR, either “soon” or possibly at launch: Astergos, Baldur’s Gate, Dota 2, Edge of Eternity, Far Cry 6, Farming Simulator 22, Forspoken, Myst, Necromunda Hired Gun, Resident Evil Village, Swordsman, and Vampire the Masquerade: Blood Hunt.
AMD’s images are difficult to check for quality, but here’s the description.
General opinions are that FSR’s “Ultra Quality” mode works pretty well, but the issues start to become more apparent as you step down through the stack. Eurogamer has some very useful comparison tools — I zoomed in on one of them to make the differences visible. While there’s only a narrow band of leaves visible, the way the bush aligns allows you to easily compare how the FSR mode impacts image quality compared with native resolution.
This continuous set of leaves makes an excellent comparison.
AMD’s 4K Ultra Quality FSR mode is very similar to native resolution. Both Eurogamer and THG note that it’s a solid model overall, and Eurogamer measured it improving 4K frame rates by no less than 42 percent. Opinions on the lower quality modes are mixed and generally get worse the lower the baseline resolution gets.
The goal is obviously for the low-end scaling models to be as good as possible, and AMD will definitely be competing against Nvidia here, but DLSS also loses a great deal of quality as you step down. One of the strengths of FSR, as compared with DLSS, is that it runs on a much wider number of cards. AMD’s RX 500 family and the RX 460 / 470 / 480 all support FSR, as do AMD Vega graphics cards, all Ryzen APUs with integrated graphics, and both the Radeon RX 5000 and RX 6000 family of products. Older Nvidia cards that don’t support DLSS, like the still-popular Pascal cards, can run FSR.
The fact that FSR runs on older GPUs could give it a practical edge over DLSS when you consider the state of the GPU market. There is some indication that GPU prices are starting to drop, but it could still take 5-6 months for prices to return to pre-shortage levels. If the crypto market heats up again or pent-up demand keeps pricing high, it could plausibly take a year or more. When massive floods put most of the world’s hard drive manufacturing capacity under several feet of water, it took two years for prices to return to pre-flood levels. Hopefully, the GPU market would correct more quickly, but that’s not guaranteed.
It’ll be easier to see comparative quality once more games support both standards (assuming any do). AMD will also presumably evolve this capability over time, just as Nvidia has evolved DLSS. Given long-term trends in gaming, it’s difficult to see them doing anything else.
When Nvidia debuted ray tracing, I wondered if there was enough headroom left in modern GPU TDPs to ramp the feature effectively. You can build specialized tensor cores and ray tracing units (Nvidia) or handle the work in your normal GPU cores and just build a lot more of those (AMD), but both companies are burning more power to handle the work.
If GPUs have to continue to add horsepower for both rasterization and ray tracing, it’s going to be even harder to offer meaningful performance improvements. But what if AI upscaling methods could one day be perfected to the point that the amount of die space dedicated to conventional rasterization could drop because no one could realistically see the difference between upscaled 1440p and native 4K?
I suspect technologies like FSR and DLSS are critical to the long-term future of AMD and Nvidia GPUs. They represent the technologies both companies hope to take advantage of in order to deliver better ray tracing in the future without compromising on other aspects of game performance. Both features will likely continue to evolve.
Right now, it looks like FSR may need a bit more time to bake, but the same was absolutely true of Nvidia and DLSS 1.0. It was the second generation of DLSS that really took a quality leap. We’re curious to see how FidelityFX Super Resolution and AMD’s other FidelityFX software packages evolve from here.
Now Read: