AMD’s Fidelity FX Super Resolution 2.0 is here, and the first title to flaunt this technique is Arkane’s wonderful Deathloop, a game that has already shown to benefit immensely from reconstruction-based upscaling. Reconstruction distinguishes FSR 2.0 from its less complex 1.0 predecessor: instead of interpolating out additional detail from the current frame with variable but often unsatisfactory results, FSR 2.0 is similar to other modern temporal supersampling techniques in that it inserts/reconstructs details from earlier frames into the current one to improve quality while adding anti-aliasing properties. FSR 2.0 is also inherently open source, which means developers can integrate it into their games for free and with minimal development time.
While we saw techniques similar to FSR 2.0 in the console space for many years, the quality is such that it deservedly takes its place among the latest second generation upscalers such as B. Temporal Super Resolution (TSR) from Unreal Engine as seen in UE5. While DLSS 1.0, checkerboard rendering, and older forms of TAA upscaling aim to produce native image quality at around half the internal resolution, FSR 2.0 and other second-generation techniques aim for similar quality at just a quarter of the resolution . with a 4K output from a 1080p base frame often cited as the sweet spot. DLSS and Intel’s upcoming XeSS use machine learning via specialized onboard hardware, but FSR uses the processing power of the GPU itself, meaning it should run on any modern graphics card. As you will see, we successfully ran FSR 2.0 on Nvidia’s venerable Radeon RX 580 and GeForce GTX 1060.
In terms of options available, FSR 2.0 offers settings that are very similar to both FSR 1.0 and DLSS. There are three modes available: Performance, Balanced and Quality. Using 4K as the output resolution, performance and quality mode uses the same internal resolution as DLSS: 1080p and 1440p respectively. Balanced mode (2259 x 1270) uses a slightly higher resolution than DLSS mode (2227 x 1253). In general, as with all image reconstruction techniques, the higher the internal resolution, the better the output quality.
So how is this quality shaped? There’s a lot to do here, and I recommend watching the video for the bigger picture. I compared FSR 2.0 to DLSS 2.3 and native resolution rendering in a number of scenarios. I have selected test cases that have historically challenged reprojection technologies involving static views, motion, sub-pixel detail, non-opaque geometry (e.g. foliage), and animation. The tests are comprehensive and worth checking out closely, although there are some screenshot comparisons on this page showing some of my work and insights.
In a nutshell, FSR 2.0 is similar to DLSS 2.3 in that it can actually look better than native resolution rendering in some scenarios: in particular, a 4K output with quality mode (internal resolution: 1440p) generally looks impressive. However, the most demanding scenarios tend to reveal more artifacts than the Nvidia counterpart. The more aggressive the chosen performance mode, the more effective these artifacts will be. However, let’s talk about it relative Quality here: Fact is, FSR 2.0 in 4K performance mode, rendered internally at just 1080p, still looks pretty impressive in isolation – a credit to AMD.
However, another caveat comes into play: quality at lower output resolutions. If you’re targeting 1080p or 1440p output, reducing the source data to work with can in turn compromise the level of detail. As a rule of thumb with DLSS, the lower the screen resolution, the higher the DLSS quality level you should use – this is even more important with FSR 2.0.
It’s also worth noting the performance considerations. As the table on this page shows, when viewed in isolation, the cost of FSR 2.0 runs faster on Nvidia GPUs than AMD hardware in almost all scenarios. In addition, DLSS 2.3 delivers more quality at lower rendering costs on RTX cards. Also, the larger the gap between the internal rendering resolution and the output resolution, the higher the processing cost of FSR 2.0. This isn’t that much of a problem because you’re saving so much more GPU time simply because you’re primarily working at a lower internal resolution – but there comes a point where, for example, pushing a lower-powered 1080p-focused GPU to output 4K via FSR 2.0 essentially becomes a waste of time.
Ultimately, the conclusions I draw from my testing are pretty simple. FSR 2.0 at 4K is fast on large GPUs and slow on smaller GPUs, but perhaps more realistically, low-end hardware performs relatively efficiently when upscaling to 1080p or 1440p resolution.
Taking a closer look at performance, FSR 2.0 was designed to offer similar performance improvements as DLSS with minimal impact on image quality. In Deathloop, for example, I noticed that the Radeon RX 6800 XT runs below 60 fps at max settings with RT enabled and native 4K. FSR 2.0 in quality mode improves frame rates by 54 percent and increases to 92 percent in performance mode, both of which enable great experiences at 60 fps (or even more). At the lower end of the hardware scale, an RX 580 outputting at native 1080p won’t be able to run the game at a consistent 60 fps on maximum settings. The older FSR 1.0 in Performance mode increases performance by 44 percent, but leaves a lot to be desired in terms of quality. Meanwhile, FSR 2.0 produces a far superior output image, although the performance boost isn’t quite as big – but a 38 percent increase is still impressive and it’s still a powerful tool for hitting 60fps.
Again, I recommend watching the video to get a better idea of the nuances of FSR 2.0, but the improvement over its predecessor is huge. FSR 1.0 provided improved results over the most basic upscalers, but performed poorly against reconstruction-based techniques up to and including DLSS. However, based on what we’ve seen so far, FSR 2.0 is a resounding success. Implementations and quality can vary from title to title – as we saw with DLSS – but AMD not only delivered a great piece of engineering, they managed to offer quality levels that are better than other software-based solutions and on par with the best. including Epic’s TSR. And if it can’t quite keep up with DLSS, then maybe it doesn’t need to – subjectively it still looks very good, and that’s about it needs to deliver since machine learning based techniques require a specific type of graphics card.
So yes, it runs on a much wider range of hardware than DLSS, so FSR 2.0 is a great thing for those who don’t have an RTX card. It’s also the first iteration of the technology, so there’s every chance that the vulnerabilities we’ve found could be improved through game patches or future versions of the algorithm. Developer adoption will be critical, but ultimately the inputs required are similar to DLSS and XeSS so adoption should be similarly rapid, and I look forward to seeing how AMD’s techniques work in future titles.