NVIDIA DLSS entered the gaming PC market as a solution to the performance deficit that the brand’s graphics cards still suffer from Ray Tracing when it is activated. It was sold at the time with a Deep Learning technology that scaled resolution upwards through AI algorithms in the company’s Saturn V supercomputer and with this it was possible to increase the perceived sharpness and the FPS, a holy grail that we already saw that it was not such a thing and that now after version 2.0 it begins to work with better perspectives not without changes.
The Ray tracing caught AMD with a different foot, since it did not seem to expect the technological deployment of its rival and in this it is late. FSR is the direct answer to DLSS, but they do not really work the same, mainly because Lisa Su’s do not have tensor units for AI and these are expected as systolic arrays in the new RDNA 3 architecture, which already have a name in a patent of the company: GSR.
But is this the solution to FSR’s problems? The displayed quality is not causing users to display good reviews, in fact, they are more “buts” than flattery as such. AMD defends itself in an interview and makes its case …
The adoption of FSR in games and software
Through the director of engineering of AMD, Nick Thibieroz, the company defends itself like a cat belly up, in an elegant and subtle way, of everything that has been flooding forums and websites around the world for 4 months:
FSR 1.0 is the result of extensive research at AMD, with several groups exploring different solutions using a variety of scaling technologies. Given the goals we had set for ourselves, we decided to release FSR 1.0 as we know it will appeal to a large number of developers and gamers who want to be able to enjoy high-quality games at faster FPS speeds on multiple platforms. without being limited by proprietary hardware.
So while I appreciate that the choice of a spatial upscaler surprised many, I think the results speak for themselves in terms of reception and adoption by developers. In fact, it has been impressive to see the various ways that FSR has been leveraged by professionals and enthusiasts thus far!
As we can see, AMD compares itself with NVIDIA indirectly and takes its chest out of the implementation of its technology in games and developers, but at the moment only 20 games have support to date, less than NVIDIA in the same period of time, which is curious after two years of development and time to spare.
FSR is not the best technique for quality
The statements continue and at one point in the interview he is referred to the fact of the graphic quality of his technology compared to that of NVIDIA. The answer leaves no room for doubt:
If you only focus on a single facet of enhancement (let’s talk about image quality), then sure enough, I think it’s fair to say that some enhancement techniques can provide better results (although in some cases you may need «pixel peeping»In still images to make this claim). I think that if we reduce the evaluation of the climbers to a single criterion, the conclusion will be incomplete.
FSR was designed to check many ‘boxes’, as we have discussed, and it is the combination of great features that makes up the complete package. Think of it like buying a new car – I don’t think anyone would base their purchase solely on how good the car looks. A savvy shopper is going to consider how fast you’re going, what options you offer, how smooth the driving experience is, and whether you can afford it in the first place.
But then why didn’t AMD use Deep Learning like NVIDIA did? Well, because it seems that ML techniques are not ideal for certain results:
Of course, if done right, ML can be a very powerful tool, but it is not the only way to solve problems. There are trade-offs you will have to make to take advantage of ML, which means you may not check some of the other really important “boxes” for a solution. Using ML in a real-time context could mean that we lose portability, performance and, if not done correctly, even some quality.
If we are objective about ML and with certain enhancement algorithms, I think the first iteration of NVIDIA DLSS is a good illustration of what I am talking about here. The mere presence of ML in a solution does not mean that you will get great results. ML clearly shows promise, and AMD is investing heavily in ML R&D on several fronts, but just because an algorithm uses ML doesn’t mean it’s the best overall solution given a set of goals.
That is to say, AMD did not take the direction of NVIDIA because it does not consider it to be a complete solution, but curiously it is working on that solution for the future and proof of this are the patents that we have been breaking down on the GSR asymmetric array units.
Not a word about the lack of quality that is obtained, or rather, the worsening of the quality of many of the games that have support, since it could be understood that they do not gain sharpness, but at least that they do not lose it and get worse the visual image. There is no answer about it, so we will have to wait as in the case of NVIDIA for a version FSR 2.0 to see if AMD can repair the problems and improve as its rival did, because version 1.0 of FidelityFX Super Resolution is following the path that DLSS did at the time.