NVIDIA DLSS 4 was one of the big stars of the keynote that the green giant held at this year’s CES, and represents one of the greatest evolutions which has experienced this technology since its presentation at the end of 2018, when the GeForce RTX 20 graphics cards were launched.
In its early days, NVIDIA DLSS was dismissed as a meaningless oddity that was destined to fail. Those who made that forecast will be surprised to see that, today, this technology is present in more than 540 games and applications, that it has a utilization rate of more than 80% among users of GeForce RTX graphics cards, and that these accumulate more than 3 billion hours played with DLSS activated.
These numbers indicate something very clear, and that is that NVIDIA DLSS has been a triumph, and not only that, it has also become one of the hallmarks of the company run by Jensen Huang, and one of its greatest values if We talk about AI and gaming. This technology has already accumulated six years of evolution, and as I said at the beginning, DLSS 4 has become one of the greatest advances in the history of this technology. Let’s find out why.
NVIDIA DLSS 4 makes the leap to transformers
DLSS technology has always used Hardware-accelerated AI. This has allowed it to improve both image quality and performance compared to conventional rescalers and frame generators, and was based on CNN models, acronym for Convolutional Neural Networks.
After six years of evolution, NVIDIA has reached the limits of CNN models, that is, it was no longer possible to continue improving DLSS with this type of models, and therefore the company has made the leap to a transformative model. This new model uses a vision transformer which is capable of performing automatic attention functions to evaluate, individually, the importance of each pixel in each frame, and also in several frames.
This type of models They use twice as many parameters as a CNN modelwhich translates into a much higher level of understanding and assimilation of the scenes. By better understanding the reality of each scene it is possible to create higher quality pixels, while also maintaining greater stability, reducing gosthing problems and improving edge smoothing. The result is a sharper image with a higher level of detail.
How this affects DLSS Super Resolution and Ray Reconstruction
The impact it has on both technologies is enormous, and in this case a picture is worth a thousand words. Look at the first image, taken from the game Alan Wake 2. As we can see, when applying Ray Reconstruction with the CNN model we have a considerable lack of sharpness in the smallest details of the fence, and all the elements of the scene look too blurred.
When using Ray Reconstruction with the vision transformer-based model The quality improvement is spectacular. The fence has an impressive level of sharpness, and all the other elements of the scene look much sharper and better defined. The difference is so big that it reminds me of the jumps that occur in games when going from low quality to maximum quality.
With Super Resolution, the difference made by the model with a vision transformer is also impressive. This new model is capable of better identifying the reality of the scene, and thanks to this improves level of detail and consistency even the smallest elements, such as hair, fur and cables.
Look, for example, at the comparison of the light cables in Alan Wake 2, with the CNN model these have small saw teeth, and the smaller ones are so blurred that they almost seem to disappear. With the model based on the vision transformer the improvement is very great, saw teeth disappear and all cables have a good consistencyeven the smallest ones.
In the second image we can see another comparison focused on the quality of the textures and their resolution. With the vision transformer-based model, the sharpness and quality of the textures is greatly improved, the blurry effect that we have with the CNN model completely disappears, and the improvement is so great that we can appreciate even the smallest details, such as the wear of the leather.
Multi generation of frames: DLSS 4 allows us to exceed 300 FPS with trajectory tracing
Frame generation was the star advancement of DLSS 3, and the great exclusive of the GeForce RTX 40. With this technology it is possible to greatly improve fluidity in games, and without having any type of dependence on the CPU. Using NVIDIA hardware-accelerated AI managed to overcome the problems and limitations of classic frame interpolationand achieved a generational leap that its main rivals have not yet been able to achieve.
With the GeForce RTX 50, the company has taken this technology to a new level, and has dubbed it multi-frame generation. To better understand why this new technology represents such a great advance, we must understand how DLSS 3 worksalso known as frame generation. This uses the information contained in two traditionally rendered frames to generate a new frame, using both data from the game engine and motion vectors.
NVIDIA DLSS 4 starts from the same idea, but instead of generating one frame it generates threewhich greatly increases the total frame rate per second, improving fluidity and maintaining high quality in each generated frame thanks to:
- The large amount of information you get from the game’s graphics engine and motion vectors.
- The highest image quality offered by the model with vision transformer applied to Super Resolution and Ray Reconstruction. Each generated frame is obtained through frames that already have both technologies applied.
- The highest performance and highest precision of the new optical flow accelerator present in the GeForce RTX 50.
Con DLSS 4 15 out of 16 pixels are generated with AI, a fact that is impressive and that only confirms the importance of this technology applied to video games. To understand its impact we only have to look at the performance data it offers, for example, in Cyberpunk 2077.
To avoid any problems with the frame generation rate, Blackwell uses what is known as “hardware flip measurement”, which transfers all the logic related to the rhythm of the frames to the display engine. This means that the GPU can more accurately manage frame display times, and it makes a huge difference, as we can see in the image.
Blackwell’s display engine has also received significant improvements over the one used by Ada Lovelace. This is capable of processing twice as many pixelsand supports higher resolutions and refresh rates.
With the game configured in 4K, maximum quality and trajectory tracing we have 27 FPS. When activating DLSS 2 the frame rate per second rises to 71, and with DLSS 3.5 we go to 142 FPS. These improvements are already fantastic, but the best is yet to come, and that is with DLSS 4 we reach 246 FPSand the PC latency goes from 65 ms in native to 35 ms with said technology activated.
NVIDIA has confirmed that the multi-generation of DLSS 4 frames will have native support in 75 games and applications from the moment of its launch, and that the new Super Resolution and Ray Reconstruction based on the vision transformation model will have native support in 50 games and applications.
For those games that have not yet made the jump to the latest version of NVIDIA DLSS, we will have the possibility of force your app through the NVIDIA app.
The multi generation of frames It only works on GeForce RTX 50. The GeForce RTX 40 will continue to use frame generation, and the GeForce RTX 30 and GeForce RTX 20 will be compatible with the new Super Resolution and Ray Reconstruction powered by the vision transformer model.
Source: www.muycomputer.com
