Quick test: DLSS 2.2 – “secret” update for better image quality in games

There is no doubt that different types of scaling techniques will arouse a great deal of interest during the year 2021. In recent years, Nvidia has spent a lot of time marketing its AI-assisted technology DLSS, while competitor AMD as recently as last week launched its own variant in the form of Fidelity FX Super Resolution. In addition, the various graphics engine makers have their own scaling techniques in place, where the goal is, as is well known, to increase performance without compromising too much on image quality.

When it comes to Nvidia’s DLSS, the technology had a very shaky launch, where version 1.0 produced, to say the least, poor results on the image quality front. However, the company did not give up and last year launched version 2.0 of DLSS, which turned out to show impressive quality results for a scaling technology. Some six months later, DLSS 2.1 was introduced, which provided support for more quality modes, dynamic resolution scaling and VR.

While DLSS in its current form produces very good results, it must be said that the technology is far from perfect. Like other temporal upscaling techniques, where information is used for previously rendered frames to build the end result, there are cases where visual artifacts can occur. How these artifacts appear varies from game to game, where these can range from barely noticeable to extremely obvious. An example of the latter is clearly seen in the image above from Death Stranding.

As with many other technologies, however, there are of course opportunities to improve DLSS with updates, where for example the machine learning model used under the hood over time can learn to handle moving objects in a better way. These improvements now appear to have been rolled out in the quiet of Nvidia in connection with Ubisoft adding support for DLSS in its title Rainbow Six Siege.

In the case of Rainbow Six Siege, curious users have discovered that the program library (nvngx_dlss.dll) used for the game’s DLSS implementation is of a newer version than previously seen in any other game. The interesting thing here is the version number, which in Rainbow Six Siege is 2.2.6.0 – something that is a noticeable starting point from previous DLSS 2.1 titles whose program library has had a version number in the style of 2.1.xx

It did not take long before people started experimenting with the new program library, and then it turned out that it was completely sonic to copy over the file to previous game implementations of DLSS 2.0 and DLSS 2.1. Interestingly, it has received reports that the new version greatly reduces artifacts in the form of lag and ghosting, which we felt was the basis for further investigation.

In this article, we thought we would take a quick look at the DLSS implementation in Death Stranding and see what graphical improvements are available from the new software library. We have copied version 2.2.6.0 of nvngx_dlss.dll from Rainbow Six Siege and compared it to version 2.1.19.0 which is part of the basic installation of Death Stranding. Of course, we also throw in a reference image without any upscaling technology enabled (native).

The tests are performed on SweClocker’s usual graphics card test system together with a Geforce RTX 3080. The resolution used is 3,840 × 2,160 pixels with the detail level “Very High”, where the screenshots have been cropped to 1,920 × 1,080 pixels for easier viewing. When DLSS is in use, the “Quality” quality mode is used, which means that the game is rendered in 2,560 × 1,440 pixels and then scaled up to the screen’s actual resolution of 3,840 × 2,160 pixels.

As can be seen in the picture with DLSS 2.1, the upscaling technique has a very difficult time handling the black particles from the game’s chiral crystals that rise towards the sky. We see quite rough-hewn tracks that follow these, where it becomes especially noticeable on particles near the camera. When we switch to the program library for DLSS 2.2, there is a marked change in the image, where the artifacts disappear almost completely – with the exception of a few duplicates.

With the two versions next to each other, the differences are very clear. The black particles really seem to make up for it for DLSS 2.1, as in addition to the tracks we also see a clear grid pattern over large parts of the sky. In addition to fixing the thorny particle tracks, this grid pattern is also blown away when we switch to DLSS 2.2 – a clear improvement!

However, the question is whether the new version of DLSS has a positive or negative effect on performance? The answer, at least in the case of Death Stranding, is that the performance is identical between the versions. For DLSS in “Quality” mode, this means in practice a frame rate gain of about 29 percent compared to if the game is run without scaling to 3,840 × 2,160 pixels.

Summary thoughts on DLSS 2.2

It’s quite interesting that Nvidia does not talk more openly about the fact that they have introduced improvements to their scaling technology DLSS. Version 2.2 of the program library seems to fix many of the major shortcomings with which the technology was drawn, where the results from Death Stranding speak for themselves. Yesterday’s video from the Youtube channel Digital Foundry also points to clear improvements in both Cyberpunk 2077 and Metro Exodus: Enhanced Edition.

Yeah, some people have seen and tried it out on different titles, and I think it’s just a testament to this generalised network being able to improve over time. It keeps on getting better, it keeps on getting more information and giving good results.

In terms of what’s new with this version, basically improvements to image quality, it just gets better. There’s nothing special about it we wanted to call out. In terms of working with developers, everyone has different cadence levels on when they take the network to put into their game, it just so happens that Rainbow Six got 2.2.

While Nvidia has not made any official announcement of the update, they have made statements on the subject, including in an interview with PC Gamer. There, the company confirms that version 2.2 provides better image quality thanks to additional training in their machine learning network, and that developers are free to implement the new library in future game updates.

Now it remains to be seen if more game developers choose to update their titles with the latest version of DLSS, as it feels a little ugly to have to overwrite .dll files in the long run. This applies not least to more multiplayer-oriented games, as such a copying maneuver can certainly trigger anti-cheating functions and get one turned off.