The Exclusive NVIDIA DLSS Tech Briefing + Demos

Spread the love



Deep Learning Using NVIDIA DNN

The first step in creating DLSS actually begins with NVIDIA generating thousands of “ground truth” reference images that are supersampled or “jittered” 64 times. Basically, the same image is rendered with 64 different sub-pixel offsets, to create an incredibly high quality 64xSS (64x supersampled) sample.

NVIDIA DLSS (Deep Learning Super Sampling) Explained!

Then NVIDIA trains a DLSS deep neural network (DNN) that uses DGX SATURNV supercomputers to generate images equivalent to the 64xSS targets. By repeatedly generating an output and comparing it to the 64xSS target, the DLSS network learns to create images that closely approximates the quality of 64xSS.

After many iterations, the DLSS network will “know” how to generate images close to the quality of 64xSS, while avoiding the blurring, disocclusion and transparency that affects traditional anti-aliasing methods like TAA. This example of DLSS 2X shows how it avoids the blurring of the semi-transparent overlay.

NVIDIA DLSS (Deep Learning Super Sampling) Explained!

After the deep neural network generates the DLSS algorithms for that game, it’s ready for use in the actual game. This is where the GeForce RTX’s Tensor Cores come in.

Tensor Cores

A key component of NVIDIA DLSS are the Tensor Cores in the GeForce RTX graphics cards. These are new processing cores designed to accelerate large matrix operations.

They allow the GeForce RTX graphics cards to deliver far superior AI inference performance than the previous Pascal CUDA cores, or a CPU for the matter.

NVIDIA DLSS (Deep Learning Super Sampling) Explained!

Using the DLSS algorithms generated specifically for that particular game, the Tensor Cores can now infer how the frame can be rendered with an image quality that is equal to, or better than, TAA with about half the shading work. Note that this is using the standard DLSS mode, and not the higher-quality DLSS 2X mode.

[adrotate group=”1″]



Here are the answers to some common questions that we believe you may be wondering about DLSS.

  1. Does NVIDIA DLSS work on Pascal or earlier GPUs?
    No. DLSS requires the use of the new Tensor Cores, that are not available in Pascal or earlier GPUs.
  2. Does NVIDIA DLSS automatically work in all games?
    No. NVIDIA DLSS requires two components to work :
    a) the game must support DLSS – NVIDIA says it’s simple for any game developer to add it to their game
    b) the graphics driver must support DLSS for that particular game – essentially, the driver should have the pre-generated DLSS algorithms for the game

NVIDIA DLSS (Deep Learning Super Sampling) Explained!

  1. How big are the DLSS algorithms?
    According to NVIDIA, the algorithms for each game are small in size – in the tens of MB.
  2. Where are the DLSS algorithms stored?
    They are stored in the NVIDIA GeForce drivers. You will need to update your driver to add algorithms for more games, as support expands.
  3. Do we need to install NVIDIA GeForce Experience?
    No, GeForce Experience is not necessary to support DLSS. You only need to install the NVIDIA GeForce driver.

NVIDIA DLSS (Deep Learning Super Sampling) Explained!

  1. Are game developers charged to have their DLSS algorithms generated?
    No, NVIDIA does the DLSS training on their Deep Neural Network for free.
  2. What’s the speed bump if I don’t use anti-aliasing at all?
    DLSS does not give you a performance boost. It only reduces the performance penalty, while delivering image quality equal to, or better than, Temporal Anti-Aliasing.
  3. What’s the difference between DLSS and DLSS 2X?
    DLSS offers image quality equal to, or better than TAA, with lower performance penalty.
    DLSS 2X offers image quality close to that of 64X supersampling – the gold standard of image quality. NVIDIA does not mention the performance penalty of DLSS 2X.

Next Page > The NVIDIA DLSS Presentation Slides, Where To Buy, Recommended Reading

Support Tech ARP!

If you like our work, you can help support our work by visiting our sponsors, participating in the Tech ARP Forums, or even donating to our fund. Any help you can render is greatly appreciated!

About The Author

Leave a Reply

%d bloggers like this: