When DLSS launched in 2019, it was a surprise hit. Although Nvidia switched from GTX to RTX to highlight how important ray tracing was going to be for its GeForce GPUs, it's actually DLSS that has become the surprise killer feature we didn't know we wanted. Of course, whenever Nvidia makes something excellent, AMD has to do the same thing, and in 2021, the company finally launched FSR, a competitor to DLSS. Here's everything you need to know about FSR, what it is, and if it's any good.
FSR: A lightweight challenger to DLSS
FidelityFX Super Resolution (FSR) is software that can enhance image quality and indirectly boost framerates, much like DLSS. It first launched in 2021 for games like Anno 1800 and Dota 2. However, one of the key differences between FSR and DLSS is that the former doesn't use AI hardware, as AMD's Radeon chips don't come with any. For the end-user, not using AI won't mean very much, but it does have some big implications for FSR and how it measures up to DLSS.
Just like how AMD has copied lots of branding schemes from its competitors, the different versions of FSR are basically analogous to DLSS's three versions. FSR 1.0 is much like the original DLSS, in that its image quality improvement is pretty meager. FSR 2, like DLSS 2, provides noticeably better image quality than its predecessor but doesn't introduce any new features. FSR 3 builds on top of FSR 2 and introduces frame generation, which algorithmically creates new frames in between the real, actually rendered frames.
One of the benefits of not requiring AI is that FSR works on pretty much any GPU from any vendor, and it has official support going back several generations, making it an option whether you have an old model or one of the latest and greatest GPUs. FSR 1 and 2 work on cards as old as the RX 500 series and GTX 10 series, and FSR 3 gets support for GPUs as old as the RX 5700 and RTX 20 series; by contrast, DLSS 1 and 2 only work on RTX-branded cards, and DLSS 3 is further limited to RTX 40.
Related to FSR 3 is AMD's Fluid Motion Frames technology, which is what makes FSR's frame generation feature possible, but it's also available standalone as part of HYPR-RX, a feature exclusive to RX 6000 and RX 7000 GPUs. HYPR-RX should work in virtually every DX11 and DX12 mode, bringing frame generation to all modern games, albeit without FSR's upscaling technology. FSR 3 and Fluid Motion Frames have just come out, and although the consensus is that it provides poor image quality and latency, that could change with updates (especially HYPR-RX's Fluid Motion Frames).
How FSR resolution upscaling and frame interpolation works, and why it's not perfect
Unlike DLSS, FSR relies on some pretty traditional image rendering techniques like anti-aliasing in order to boost image quality, but combines it with an algorithm that wrings out a little more fidelity. There's no performance-boosting going on here despite the fact FSR is billed as being capable of it. Instead of using FSR to make things look better at a resolution, AMD uses it to make a lower resolution look like a higher resolution, and that's how it boosts performance. 720p with FSR enabled to bring it to 1080p isn't actually 1080p, but a lower resolution that looks close to 1080p.
Here are the basics. FSR 1 would first spatially upscale a real rendered frame and then sharpen it. It wasn't anything fancy, and it also relied on games to provide good anti-aliasing, so the actual frame the GPU renders has the best shot of getting upscaled properly. FSR 2, on the other hand, is much more complex, replacing anti-aliasing while adding a bunch of other features like depth and motion vector buffers. Essentially, FSR 2 gives a much bigger makeover to rendered frames than FSR 1, which makes FSR 2 even better at increasing image quality.
FSR 3 meanwhile is pretty similar to DLSS 3 frame generation; it just uses a regular algorithm instead of dedicated AI hardware too make it work. While running rendered frames through FSR 2's image quality boosting algorithm, FSR 3 can then take the difference between two different, actually rendered frames and create a new, generated frame in between. Enabling FSR frame generation can boost the framerate by at least 50% and in some cases nearly 100% or double.
Instead of using FSR to make things look better at a resolution, AMD uses it to make a lower resolution look like a higher resolution, and that's how it boosts performance.
Of course, FSR shares pretty much all of DLSS's inherent problems. Firstly, FSR 1 and 2 won't do anything if you run into a CPU bottleneck. If your CPU is bottlenecked, whether it's because your CPU is too slow or old or because the game is poorly coded, it means you can't really boost the framerate by lowering settings because these settings almost exclusively reduce the load on the GPU rather than the CPU. Resolution is a purely GPU-dependent graphics option, and lowering it as FSR 1/2 does won't boost the framerate if there's a CPU bottleneck, which can only be realistically resolved by having a very good CPU.
As for FSR 3, it has the same issues as DLSS 3 based on early reviews. It has to replicate literally everything you'd normally see in a finished frame, including UI elements, and FSR 3 isn't good at that, even worse than DLSS 3. The other problem is latency. Since you need two frames in order to make one in between, you'll see frame two (the interpolated/generated one) after frame three has already been rendered, which means you're reacting to the game many milliseconds slower than usual. It's almost as if you're playing at a lower framerate and not a higher one, though you are visually seeing lots of frames.
It is technically possible that the visual quality issues of FSR 3 could be improved in the future. In the case of latency, it's literally impossible to reliably create a frame with only a single reference unless FSR 3 can somehow predict the future, so latency will be higher no matter what, though it's possible AMD could whittle the latency down from bad to tolerable.
FSR still isn't quite as good as DLSS, but it has one big advantage
The consensus among reviewers is that while FSR is good, Nvidia's technology is generally better. It's in more games and has higher visual quality, which is all that really matters to the end user. However, FSR has two big advantages DLSS can't really match.
The first is compatibility. FSR 1 and 2 work on pretty much any GPU, and FSR 3's level of support is good too. Although Nvidia has been making RTX cards since 2018, lots of people are still using older GTX-branded cards from the 10 and 16 series. Those users can't use DLSS because their cards don't have Tensor cores, but they can use FSR 1 and 2. It might seem ironic that Nvidia (and Intel) users stand to benefit from FSR as much as owners of AMD cards, but it gives FSR lots of utility.
The rate of FSR's adoption is also very significant. Although DLSS had nearly a three-year head start on FSR, both technologies are in a similar amount of games. At the time of writing, it seems that around 300 or so games have DLSS in one form or another, while FSR was present in 230 games as of December 2022, so we can assume maybe 250 games today have FSR 1/2. That's not a bad level of support by any means for FSR, and it also means FSR is getting added to more games faster than DLSS.
This is probably because it's easy to implement in games, at least according to AMD. In games that lack all the technologies FSR 2 requires in order to work, it takes four or so weeks to get FSR 2 up in running, but it can take as little as three days to add FSR 2 to a game that already has DLSS 2. By contrast, developers need to work directly with Nvidia for DLSS support. FSR is open source, and it's up to the devs to integrate it into games. Of course, the obvious problem is that there are only so many games with DLSS 2, so FSR's momentum could certainly slow down if it hasn't already.