Does the new technology stand up to DLSS?
AMD introduced along with the launch of Radeon RX 6X50 XT series the debut of version 2.0 of FidelityFX Super Resolution technology, AMD FSR 2.0. This technology has an objective analogous to that of Nvidia Deep Learning Super Sampling (Nvidia DLSS), and the latest Intel Xe Super Sampling (Intel XeSS): to make it possible to render the game at a lower resolution, use an image enlargement technique while keeping as much of the details, and thus deliver more performance with low loss of visual quality.
More than an evolution, FSR 2.0 is a totally different technique than FSR 1.0
AMD FidelityFX Super Sampling 2.0 is AMD’s own “built from scratch” technology, so different from version 1.0 it almost deserves a name of its own. Let’s unravel AMD’s new strategy, the result in performance and graphics quality and also how the dispute with Nvidia DLSS is. Do we no longer need tensor cores and, consequently, RTX, to do a good upscaling?
Really an AMD DLSS
The first big change from AMD FSR 2.0 is that it now really closely resembles Nvidia’s approach to DLSS, or rather DLSS 2.0. AMD FidelityFX Super Sampling 1.0 did the same as all the technologies listed so far, but with a different technique. He would take an image, enlarge it, and use FidelityFX Constrast Adaptive Sharpening (FidelityFX CAS) to maintain image detail and definition.
The new approach is closer to DLSS
AMD FSR 2.0 now uses more information, with a temporal filter. That is: in deciding how each pixel will look, FSR 2.0 takes into account information from previous frames. It will also use other information from the game engine, the motion vectorsor motion vectors, which will inform this technology where things are and where they are going, helping again to make the right choices for resizing the image.
If you’re experiencing deja vu, you’re not wrong, because you’re probably remembering our article about DLSS version 2.0 and how well doing it worked for the super sampling technology on the GeForce RTX side.
FSR 2.0 takes motion vector information, object depth information and colors, mixes it with data from previous frames, and makes a decision on how to fill in the missing spaces to increase image resolution.
The approach of FSR is very similar to that of DLSS in its modes. Essentially you’re downsampling the image is actually rendered, and you’re using more machine learning to fill in the gaps in higher performance modes. Below is the progression of these modes, remembering that they reference the final resolution, or output resolution, in this table:
– Continues after advertising –
Comparison DLSS vs FSR 2.0 in video
FSR 1.0 vs FSR 2.0 and vs DLSS graphics
Time to see the evolution versus FSR 1.0, and how the dispute with DLSS is in terms of graphics. We made comparisons with 4K resolutions and also Full HD, with entry and intermediate cards from both Nvidia and AMD. For the first test, let’s see the performance in a 4K scene:
Below we will analyze the quality progression of the FSR in its performance, balanced and quality modes in this scene. To make the difference clearer, we are using a 5x zoom, to capture the details:
The result of the quality mode is excellent. It’s evident how it manages to retain a lot of detail, and thanks to the sharpening effect we get the feeling that the resolution is even higher than the image actually rendered at 1080p. Turning down to performance mode, the image starts to deteriorate, especially showing more artifacts on moving objects, something that becomes more evident if you watch the comparison video. Despite the loss of quality, it’s still quite viable to play FSR 2.0 in performance mode, especially when you’re not with the “nose glued to the screen” that is the 5x zoom effect.
Now let’s go to the comparison again, but putting the FSR 2.0 to be compared with the 1.0 and also the Nvidia DLSS:
– Continues after advertising –
Putting into action the 4K Performance mode of the three technologies, that is, all are rendering in Full HD and upscaling to 4K, the evolution of FSR 2.0 is remarkable. The image has much more definition, retains more detail and has less artifacts at the edges of moving elements. DLSS still has an advantage, with more defined edges, less artifacts on objects, but the difference is already negligible, especially without the zoom we used in the comparison above.
Let’s move on to another scene, now to investigate the differences between these technologies in their quality modes, that is, starting from 2560×1440 to 3840×2106.
We have two zoomed segments in this scene:
In this part of the comparison, the static images do not do justice to the elements that are more evident in the video. Both native 4K with TAA and FSR 1.0 have serious flickering issues at the edges of objects. FSR 2.0 and DLSS reduce or even completely correct most of them, delivering a much more finished and better image than native 4K rendering. DLSS has the advantage and handles some patterns better, such as the bars stuck on the roof of the building, but in this comparison FSR 2.0 better solved the character’s movement in the background. Here I recommend again seeing this scene compared on video.
Now let’s see what FSR 2.0 technology looks like at a lower resolution. In this case, we go with a Radeon RX 6600 handling the game in Ultra, RT On and Full HD resolution.
Again, the flickering is an element that will only be visible in the video comparison, but I already emphasize that it was evident how much FSR 1.0, even in Quality mode, suffers a lot from this problem. FSR 2.0 in Performance mode also had this issue, in addition to losing definition to a similar degree to what happens in FSR 1.0 Quality. In short: for 1080p the ideal is to aim for FSR 2.0 Quality mode. The graphical result is very good, and although the performance gain is modest, it is more interesting than running natively and applying TAA.
This turns out to be an unavoidable effect of working at a very low resolution. Like DLSS, FSR 2.0 suffers when the final resolution is too low. In quality mode, it renders from HD to deliver Full HD, and in performance mode we have a measly 890×540 as the base resolution, which is little information for the upscaling reconstruction.
FSR 2.0 performance
We compared the performance of FSR 2.0 with the 1.0 version of the technology and also with the DLSS, which in Deathloop is implemented in the 2.3 version of the Nvidia technology.
Entrance and old signs
Starting the comparisons with the older cards, we see that the popular Radeon RX 580 and GTX 1060 6GB take so much benefit in performance, especially the Radeon. Going to the latest cards, like the GTX 1650 GDDR6 and Radeon RX 6500 XT, we get two opposite results.
The GeForce GTX 1650 GDDR6 took a lot of benefit from the FSR 2.0, with a gain of 20%, while the Radeon RX 6500 XT practically stalled at a pretty bad frame rate. In fact, this card ran all tests poorly, showing that the limitations of this model – which we’ve already got on our feet in the past – may have negatively influenced this test.
AMD FSR 1.0 still delivers a greater performance gain, but as we show in the comparisons, it’s not on the same level as the other two technologies.
Starting with Full HD cards, like the RTX 3050 and Radeon RX 6600, we have good results enabling FSR 2.0 Quality, and interestingly, the GeForces took the most advantage. The RTX 3050 has increased its performance by almost 25%, with a gain of 15% which is still good, but much less relevant. DLSS fared slightly better than FSR 2.0, while FSR 1.0 and its simpler implementation remains the best test result.
And finally, we took a look at 4K with the RTX 3060, a card that doesn’t have the performance to handle the game at that resolution. All technologies have given a respectable boost, with FSR 2.0 already bringing 30% gain and putting the average at 50fps. This shows that just using the FSR in Performance mode, which still has a good level of quality, or changing the preset from Ultra to High, to make 4K gameplay viable on the RTX 3060. But it’s even more game to take advantage of the tensor cores and enable DLSS, which will bring gameplay to 60fps+ in High quality.
AMD FSR 2.0 still has some slight drawbacks compared to DLSS 2.3, but it wouldn’t even need to tie or beat Nvidia’s resource. Just get close enough to show itself as a valid alternative for Radeon owners. And he did.
AMD FSR 2.0 didn’t even need to beat DLSS, just get close enough. and he got it
Version 2.0 of FSR is comparable in quality to DLSS, and if it weren’t for our zoomed-in reviews, it would be hard to tell the difference. With gameplay going on, it’s unlikely that anyone will see the difference, unless you stick the pixels under a microscope and stare. This is great news not only for AMD card owners, but also for GeForces owners without DLSS support.
But it’s not all breakthrough, and AMD FSR 2.0 needs a lot more developer intervention. Fortunately, DLSS 2.3 paved the way for this kind of temporal filter with upscaling reconstruction, so possibly adoption will be faster than DLSS in its early days. But it is still something to be seen in practice.
Owners of DLSS capable cards should continue to take advantage of the tensor cores of their GeForce RTXs. DLSS is faster than AMD FSR, delivering more frames and having slightly better graphics results. Next, AMD FSR is a great option, and should be used in 4K even in Performance mode, while Full HD is ideal to use in Quality mode. If that doesn’t deliver enough frames to play, the way is to go with FSR 1.0, which is even more interesting than not using the feature, in many cases it might be better to stick with native resolution in these cases.
Source: AMD Community, GPU Open