In the field of technologies that improve the number of frames per second without (too) degrading the image quality, DLSS, from Nvidia, is both the oldest technology, but also the most complete and the most mature. .
But its competitors Intel and AMD are also developing similar technologies. Like AMD’s FidelityFX Super Resolution – FSR –, which will soon be in version 2.0 as the company has just announced.
On the same principle as the DLSS, the FSR calculates an image with less definition than that of the target screen, and puts a magic wand on it to properly enlarge this image – we speak of scaling or “ upscaling” in the jargon.
In version 1.0, FSR was already more open than DLSS, since it works on all GPUs – yes, even those of AMD’s competitors. But it was much less efficient, especially in terms of image quality. Where Nvidia and game developers build title-specific DLSS models, the FSR 1.0 was less bespoke and more basic.
Read also: Nvidia improves its AI which makes games even more beautiful (March 2020)
FSR 2.0 will remain a technology open to all GPUs, but it becomes more selective in terms of titles. It now requires a game-specific implementation for this toolkit to be active – we don’t know if this is a trivial implementation or if, like DLSS, it requires a lot of time and of sweat.
Vectors to the rescue, but still no ML
The big novelty brought by FSR 2.0 is the addition of so-called temporal data – basically, vectors that characterize the movement of pixels. While FSR 1.0 focused on enhancing each image independently – spatial data: magnification filters are applied to each frame – FSR 2.0 uses vectors to “track” pixels. This allows calculations to be anticipated and improves quality (no need for Temporal Anti-Aliasing – TAA – for example).
However, due to its universal nature, FSR 2.0 does not use any neural network to improve image quality – the strong point of Nvidia’s solution.
Logical in a sense: no Radeon card has this type of calculation unit. While image quality from one FSR generation to the next should improve, DLSS and Nvidia’s chips — which incorporate machine learning (ML) compute units — still, on paper, have l advantage in speed and accuracy. so quality.
It therefore remains to wait for the arrival of FSR 2.0 during the next quarter, and to characterize, supporting games, the performance gains of each solution. As well as the differences in image quality, the difficulty being to determine, title by title, if they are really noticeable.
For now, the only demonstration is the YouTube video above showing the implementation of FDR 2.0 in the game death loop of the French from Arkane. Let’s hope that AMD quickly produces beta versions as well as downloadable files (without applied compression) to analyze the results with a magnifying glass.
In any case, between the DLSS, from Nvidia, the FSR, from AMD, and the XeSS, from Intel (Xe Super Sampling), it seems clear that these upscaling “tricks” are, like the reduction of the fineness of engraving, a major avenue for improving chip performance.