WHAT IS NVIDIA'S DLSS? GUIDE TO THE AI OF WONDERS
The challenge between NVIDIA and AMD, the only players in the video card market for some time now, has seen a slow but inexorable change in the battleground. The comparison is no longer limited to just the muscle level, with faster and quieter cards, capable of supporting ever higher framerates and resolutions, but is moving more and more on unique features capable of upsetting the market.
NVIDIA with its DLSS aims precisely at this, and in the next
lines we will analyze its operation, as well as the reasons that make this
technology a real "game changer" .
The wonders of AI
At the same time as the releases of its Turing
architecture-based video cards, the RTX 2000, NVIDIA unveiled the first version
of its neural network-based super sampling, DLSS (Deep Learning Super
Sampling). NVIDIA's goal is to undermine the paradigm that has always bound
graphics improvements on PCs to the brute power of video cards, bypassing the
problem in a completely innovative way.
In fact, DLSS allows to render images at lower resolutions
and to have them reconstructed, at higher resolutions, by an artificial
intelligence , so as to have to process a much lower number of pixels without
compromising on graphics and performance.
This technological miracle is possible thanks to the Tensor
Cores, calculation units present in the Turing and Ampere chips and highly
specialized in accelerating calculations in FP16 and FP32 in 4x4 matrix, so as
to be able to generate a very high number of floating point operations per
cycle (285 Tflops on the RTX 3090), precisely those necessary for all
activities related to deep learning and inference.
Before the introduction of Tensor Cores it was unthinkable
to hypothesize image reconstruction techniques based on AI (by the way, in this
article we deepen the best AI in video games ) and processed in real time on
consumer video cards.
The first version of DLSS needed to be "trained"
in a totally independent way for each individual game. NVIDIA fed its Saturn V
DGX supercomputer clusters a multitude of reference images (rendered at very
high resolutions, up to 64x supersampling) in order to generate a
"perfect" one. Later after comparing it to a normally rendered image,
it prepared its AI to reconstruct the missing details.
Through a process of backward propagation of the error, the
AI trained itself to generate an image very similar to the "perfect"
sampleand devoid of the typical flaws that plague many other anti-aliasing and
upscaling techniques, such as blur. The idea behind the technology was already
excellent but the results obtained were not always excellent with the first
version of the DLSS.
With DLSS 2.0 NVIDIA has taken a further step forward, the
learning process is no longer specific to each individual software but is based
on a generalized network that makes implementation much simpler and more
effective, no more thousands of images are needed. in very high resolution to
reconstruct the scene, the AI is powered by a combination of low resolution
images and motion vectors generated by the same game engine. Vectors are
responsible for predicting which direction objects are moving in the game and
are used to understand what the next frame will look like.
The AI network on which DLSS 2.0 is based, through a
convolutional autoencoder, then takes care of reconstructing the image at a
higher resolution.. At the end of this training phases the image will be
compared with a rendered 8640p to verify the goodness of the results obtained
from the neural network processing. Once this process is completed, the
generated model is integrated into the Game Ready Drivers and made available to
all owners of the RTX series video cards. The benefits are incredible: being
able to render up to a quarter of the pixels compared to the chosen resolution
(based on the selected quality setting) facilitates the achievement of very
high resolutions such as 8K, allows you to invest the saved power in effects
and increased framerate and makes it possible to activate features such as Ray
Tracing, which would otherwise have a very important impact on performance.
The most interesting aspect of this technology is that in
many cases the artificial intelligence not only increases the performance, but
also improves the overall yield compared to a natively rendered image, which is
unthinkable with any other upscaling technique.
DLSS Gaming
Currently the DLSS is supported by a good number of titles,
a striking example of the goodness of the solutions implemented by NVIDIA are
the results obtained on Control . The frenetic action of Remedy immediately
used all the new technologies introduced with the RTX cards such as DLSS and
Ray Tracing. In its first version (DLSS 1.0) the upscaling technique, although
it allowed the game important performance gains, compromised its overall
performance by adding some graphic artifacts and a perceptible blur effect on
the whole scene. It is with the support of DLSS 2.0 that the situation has been
totally turned upside down.
Introduced at the same time as "The Foundation"
update, the new version of the AI-based upscaler has maintained the performance
leap recorded with DLSS 1.0, but resolved all those graphic defects that bored
the first release . The result is extraordinary, more information is available
in our special dedicated to Control .
The discussed Cyberpunk 2077 sports - especially on PC -
truly exceptional graphics: Lighting in Ray Tracing, volumetric fogs and
breathtaking draw distances are just a part of the graphic wonders proposed by
CD Projekt RED, which certainly make the title a joy for the eyes, but which
heavily impact the performance of the game, making the DLSS in fact the only
admission ticket to the highest settings.
Even Death Stranding on the occasion of the release on PC
which took place last summer, has immediately integrated the technology of
NVIDIA. The result is also incredible in this case, the image is much sharper
even when compared to its counterpart natively rendered in 4K., the writings on
the packages that Sam carries across the wide valleys of Hideo Kojima's
post-apocalyptic are more readable after processing the AI from NVIDIA. In
addition to this, a performance gain is obtained that also allows owners of an
RTX 2060, the lowest entry point to access the wonders of DLSS, to push the
title up to 4K while maintaining 60 fps.
Configure the DLSS
Starting with the second release of the DLSS NVIDIA has
decided to simplify the settings of its upscaler to make it as user friendly
and easy to configure as possible. Almost all titles share three settings:
quality, balanced and performance, with some exceptions such as Cyberpunk 2077,
which adopting the latest version of DLSS (2.1) also allows you to select an
ultra performance mode . The only difference that separates the three possible
settings is the native resolution from which the upscaler will work its magic.
Control, unlike other titles that use DLSS, allows us to
directly choose the resolution of the native rendering and, despite being a
less intuitive solution, it gives us the opportunity to take a quick look at
the inner workings of the upscaler. On a 1440p monitor, for example, the
possible settings are 1280x720, 1484x835, and 1706x960, and correspond to the
performance, balanced and quality settings.
To make the comparison even clearer with a 4K monitor the
lowest possible resolution is 1080p, which means that 3 out of 4 pixels on the
screen are computed totally by NVIDIA AI ... quite revolutionary huh?
So what are the recommended settings for DLSS? Difficult to
give a univocal answer, obviously the lower the quality level the more the AI
will struggle to make the image sharp as at native resolution, but there are
several factors to keep in mind. First of all, the resolution and the diagonal
of the monitor are very important, certainly monitors with high resolutions and
low diagonals allow much more aggressive settings, while on monitors with large
diagonals and low resolutions it is not advisable to go below the quality
setting.
Another factor not to forget is that the AI, as such, will
improve over time and probably with future revisions of the NVIDIA drivers or
the DLSS technology itself we will have important improvements. Our advice is
to always try the three settings, and find the right balance between
performance and graphic fidelity.
The future of DLSS
On balance, the concept of DLSS draws an important line and
lays the foundations for a future of gaming increasingly dependent on
artificial intelligence, both in the development phase and in the rendering
phase. The innovations introduced with DLSS 2.1, such as support for VR titles,
8K gaming and dynamic resolutions leave us hoping for an increasingly rapid
spread and we are very curious to see the future developments of this
technology. AMD for its part has not stood by and announced that it is working
on a technology similar to DLSS , which will be open and cross-platform, and
which perhaps can also be used on Series X and PS5.
