Nvidia DLSS on Turing in the test
: Test |CUP | Specs |Config
tl;dr: With DLSS, Nvidia was unable to keep its great promise for a long time. The first games ran faster thanks to the reduced rendering resolution, but the quality didn’t drop. With DLSS 2.0, AI upsampling is now taking a decisive step. But the technology is not there yet.
Nvidia’s Turing graphics cards come with numerous new features. Without a doubt the most important is ray tracing, which will also be available in the future with RDNA2 on AMD graphics cards and the next-gen consoles PlayStation 5 and Xbox Series X. Raytracing uses the RT cores at Nvidia.
However, DLSS (“Deep Learning Super Sampling”), which is supposed to use the tensor cores of the architecture, was also highly praised for the presentation. This is an intelligent upsampling based on an AI algorithm. DLSS is supposed to make it possible to render a game in a lower resolution without losing image quality. The AI algorithm is said to have been trained beforehand in a neural network at Nvidia for an optimal result.
Top in theory, solala in practice
In theory, DLSS sounds good. In practice, however, the technology has worked at best mediocre to poorly so far. Final Fantasy XV looks quite alright with DLSS, while Metro: Exodus created quite a bit of mud, especially at the start, and Control was a mixture of blurry images with graphic errors. Even if image quality is always a matter of taste and many critics overturned, the editors and many readers in the ComputerBase forum agreed: DLSS did not deliver the promised result.
But Nvidia stuck with it – on the level of the games and in the background. Contrary to the official PR line, Nvidia was probably not yet satisfied either. And after a number of 1.x derivatives, the time had come with DLSS 2.0: DLSS should be significantly better in the new generation.
ComputerBase took this as an opportunity to take another look at the “intelligent upsampling” and check whether it is now really possible to gain performance in a meaningful way.
DLSS 2.0 unites a network for all games
DLSS 2.0 comes with numerous changes compared to DLSS 1.x. It starts with the neural network itself. While up until now every game had to be trained separately by the AI algorithm, with DLSS 2.0 there is a single network that can be applied to all games. As a result, Nvidia hopes, among other things, a significantly faster spread of the still new technology. The neural network now not only works universally, but should also have become significantly faster, so that the power that has become free can be put into a higher image quality.
Tensor cores are only now being used
In line with this, DLSS 2.0 is now using the Turing or GeForce RTX-exclusive tensor cores for the matrix calculations for the first time. Because even if it has always given the impression that DLSS 1.x has already been calculated on the tensor cores, this was not the case. Instead, DLSS has been running on the classic shader units until now. With DLSS 2.0, on the other hand, the FP16 calculations run via the tensor cores, so that the classic ALUs can take care of the graphics alone.
Three quality levels and more flexibility
DLSS 2.0 is significantly more flexible than the initial versions. Up until now the games had to struggle with the problem that every graphics card could only use DLSS with certain resolutions. For the entry-level GeForce RTX 2060, for example, DLSS was only available in FHD, whereas DLSS in higher resolutions was only possible on faster models. DLSS 2.0 now does away with this and can be used in all resolutions.
DLSS 2.0 offers three different profiles in all supported games: “Quality”, “Balanced” and “Performance”. These affect the rendered resolution and thus both the speed and the image quality. Control is the exception, because there three “DLSS resolutions” can be set explicitly for each resolution. Presumably, these are also the three profiles, only without a name, but with an indication of the rendering resolution. In addition, it can be assumed that the remaining DLSS 2.0 games use the same resolutions as Control, only under a name instead of the number of pixels.
With the changes, Nvidia wants to ensure that DLSS 2.0 achieves the same quality on average as the native resolution – with significantly better performance at the same time.
Test system and games
The following graphics card benchmarks were carried out on an Intel Core i9-9900K, which is operated with the standard settings and can access a total of 32 gigabytes of RAM at a speed of DDR4-3200 (14-14-14-32). “Windows 10 November 2019 Update” (1909) including all currently available patches is installed. The GeForce 442.75 was used as the driver.
Control, Deliver Us The Moon and Wolfenstein: Youngblood were used in the games. In addition to DLSS 2.0, the three games support ray tracing, which was activated throughout the test. In the three titles, DLSS is tested in Full HD and Ultra HD in all DLSS quality levels. Graphics details have been maximized in all games.
DLSS 2.0 looks a lot better than it did in the beginning
DLSS has not yet been able to convince in any game. Not even in Control, which with DLSS 1.9 offered a very advanced evolutionary stage of the 1st generation. But the game still offered a lot of blurring with the AI upscaling. And the many graphical errors that Control has were often made worse by DLSS 1.9. The better performance didn’t help either. With DLSS 2.0 the tide has turned there too.
DLSS 2.0 in Deliver Us The Moon
The graphics of Deliver Us The Moon are generally not really sharp, and there is intentionally good-to-see noise in the image. And DLSS 2.0 does that properly in the quality setting and thus a still relatively high rendering resolution. So neat that DLSS looks better in some scenes than with the native rendering resolution, but in others details are lost. It is noticeable how fluctuating the quality is, but on average both settings do not give each other much. DLSS also has the advantage that in Full HD some objects are rendered better at a greater distance, some of which disappear without intelligent upsampling. A clear winner cannot be chosen in the game.
Deliver Us The Moon – Full HD, Nativ
image 1 from 8
The more aggressive DLSS modes can also be used surprisingly well in Deliver Us The Moon. Especially in Ultra HD you have to be very careful to see a difference between the quality and the performance setting. Full HD suffers more from the performance setting, but still leaves a usable impression. Only with very fine objects such as a fence is the internally rendered low number of pixels no longer sufficient and the image begins to flicker.
Deliver Us The Moon – Full HD, Nativ
image 1 from 8
DLSS 2.0 in Wolfenstein: Youngblood
In Wolfenstein: Youngblood, DLSS 2.0 has two sides. The very positive one has to do with the game as such. The offshoot of the Wolfenstein series has to struggle with a very blurred edge smoothing, which creates an annoying blurring even in high resolutions. Because DLSS 2.0 does the anti-aliasing itself and switches off the game’s own anti-aliasing, there is no problem – with the result that Wolfenstein: Youngblood with DLSS 2.0 looks almost consistently sharper and therefore better in the quality setting. This even applies to Ultra HD, but above all to Full HD, which usually looks much sharper with DLSS 2.0 and also flickers less when moving. Because the game’s anti-aliasing is not only fuzzy, it is also not very smooth.
Wolfenstein: Youngblood – Full HD, Nativ
image 1 from 8
Graphic errors tarnish the otherwise very good impression
However, there is also a second page, because without playing too long or explicitly looking for errors, there are, of all places, in the headquarters, to which you keep coming back, to see sometimes massive graphic errors with DLSS (a wide band from left to right) . Even different resolutions or graphics settings do not change anything in the difficult-to-describe errors. This is doubly annoying, because without them Wolfenstein: Youngblood with DLSS 2.0 would have been a nicer game than without the AI upsampling. Nvidia has been informed about the problem for a few weeks, but has not yet commented on it.
Wolfenstein: Youngblood – Full HD, Nativ
image 1 from 8
The lower DLSS modes visibly lose their sharpness compared to the quality setting, but even DLSS on “Performance” does not look any blurrier than the native resolution on average. Some objects show less detail, while others still show more. In addition, DLSS still has the advantage of better anti-aliasing. On the other hand, the graphic errors in the headquarters increase even further and literally jump in the face.
DLSS 2.0 in Control
DLSS 2.0 also has its strengths and weaknesses in Control. It is absolutely positive that the game visibly loses its extremely pronounced blurring regardless of the resolution and the DLSS level. That affects the whole picture. Many details can now be seen much better. The subtleties on surfaces only become really visible with DLSS 2.0, switchboards can be seen much better and the texts on posters can be read really well for the first time. Even in Ultra HD, DLSS 2.0 is significantly sharper than the native resolution, but the same applies to Full HD.
Control – Full HD, Nativ
image 1 from 8
If it stayed that way, Control would be a much nicer game thanks to DLSS 2.0 despite noticeably better performance. But there are also some problems – significantly fewer than with DLSS 1.9, but still sufficient. With DLSS 2.0, the anti-aliasing does not come close to the quality of the native resolution, even in 3840 × 2160. This may be the result of the significantly better sharpness, but the game flickers a bit more.
That alone is not serious, but DLSS 2.0 still likes to jump on the graphics errors of the game itself and make them significantly worse. Then some objects are completely pixelated, which is particularly noticeable when displaying grids – and there are a lot of them in Control. In addition, it occasionally happens that the AI algorithm has difficulties with denoising for ray tracing. Then some reflections continue to flicker even without movement.
Control – Full HD, Nativ
image 1 from 8
The problems occur in Ultra HD and are much more pronounced in Full HD. DLSS is probably not responsible for anything but the problematic graphics of the game itself, but with the AI upscaling, the difficulties are much more significant. The same applies to the DLSS levels “Balanced” and “Performance”, which are even more prone to problems than the “Quality” setting. On the other hand, the significantly better image sharpness is also retained in the performance setting.
On the next page: Benchmarks with DLSS in Full HD and UHD as well as the conclusion