DLSS 3 is an evolution of the well known scaling and reconstruction technology supported by artificial intelligence from NVIDIA, which maintains this base, but at the same time introduces a very important innovation that allowed it to make a significant leap in performance without depending on other components: frame generation.

DLSS 3 technology has two parts, one is the scaling and intelligent image reconstruction process that was already present in DLSS 2, and the other is image generation by artificial intelligence. The latter represents the important innovation we talked about and will delve into in this article.

However, we will not limit ourselves to frame generation only. To better understand what DLSS 3 is, how it works and what value it brings We will look at its two parts in depth. However, if you have any questions after reading this article, you can leave them in the comments and we will help you solve them. Without further ado, let’s get down to business.

NVIDIA DLSS 3: Image reconstruction and scaling

DLSS 3 technology preserves, as we have already said, the process of image reconstruction and scaling. Thanks to artificial intelligence, the algorithms used by this technology can create images with a high final resolution and high image quality, starting with images rendered with a lower pixel count to the target resolution.

The process is very easy to understand with an example. When we use DLSS 3 in performance mode and have a target resolution of 3840 x 2160 pixels, the render resolution will be 1920 x 1080 pixels. algorithms combine the previous images with the current imageand resort to spatial and temporal elements as well as motion vectors, create a high-quality frame that preserves excellent detail in both near and far objects and even those with low geometric loading.

Traditional upscaling systems don’t use artificial intelligenceand they don’t even resort to temporary elements, just are limited to padding pixels starting with a spatial basis, which in the simplest cases is equivalent to “pixel stretching”. Other more advanced ones use fixed algorithms that slightly improve the result but do not reach the level we would get with DLSS 3 because they tend to they show a much lower resolution and lower level of detail, which ultimately results in a blurry and bland appearance. They are also prone to issues with flickering and fading of distant objects, especially when they have a low geometry load.

Thanks to the image reconstruction and scaling process offered by DLSS 3, which is also present in DLSS 2, it is possible to create an image with a quality very close to native 4K already from 1080p resolution. It follows significant performance improvement, and minimal sacrifices in terms of image quality. However, reducing the resolution increases the dependence of the graphics card on the CPU, so reducing the resolution to a level where the processor can end up creating a bottleneck.

Frame generation: solving the CPU bottleneck

When DLSS 3 was introduced by NVIDIA, they confirmed that the most important innovation that this technology introduced was frame generationwith this working in conjunction with image reconstruction and scaling to further improve performance.

The process of image scaling and reconstruction using artificial intelligence takes place in tensor coreswhile frame generation runs on “Optical Flow Accelerator”component found on GeForce RTX 20, GeForce RTX 30 and GeForce RTX 40 graphics cards.

This graphics engine accelerate the entire workload which includes frame generation and works with AI DLSS Frame Generationwhich is the one that decides how all the information obtained during the frame generation process is used and the one that defines the final result.

Frame generation is intended to improve performance but it does not resort to downscaling or optimizing the reconstruction and scaling process. This technology analyzes two sequential images that were rendered in the traditional way, that is, with CPU intervention, and uses information from these images and previous images to generate another image completely independent, that is, without CPU intervention. CPU, CPU.

We might think that DLSS 3 “invents” the frame, but that would be a mistake, because what it actually does is generate another frame with the information it gets from other frames and uses artificial intelligence, motion vectors and optical flow (“Optical Flow”) get an accurate prediction of how all the objects and elements in that frame should look, including things as complicated as shading and lighting.

With this technology we have one extra frame for every two traditionally generated frames. Said frame is not affected by the CPU as it is completely independent as we mentioned, and thanks to the use of motion vectors, optical flow and intelligent frame analysis, DLSS 3 manages to maintain very good image quality.

The performance improvement we can achieve with frame generation is so great that depending on the game, this can increase by 548% performance vs. native mode. For comparison purposes, the improvement we would have achieved with just DLSS 2 in the same scenario would be 329%.

However, frame generation comes at a cost in terms of latency. This can be a problem in specific cases, but NVIDIA has solved it with reflective technologywhich automatically activates when it detects that framing is working, and performs a synchronization task between CPU and GPU which can significantly reduce latency and make the response speed in games good enough for you to enjoy an optimal experience.

DLSS 3 compatible graphics cards and supported games

The center frame was created using DLSS 3

We already said that DLSS 3 uses tensor cores and an “Optical Flow Accelerator” and we know that image generation is done on top of the latter. GeForce RTX 20, GeForce RTX 30 and GeForce RTX 40 have tensor cores and also equipped with “Optical Flow Accelerator”, but only the latter are supported for frame generation.

This has an explanation, namely that the “Optical Flow Accelerator” present in GeForce RTX 20 and GeForce RTX 30 It does not have the computing power or sufficient accuracy, how to get a good result with image generation. For this reason, NVIDIA has limited DLSS 3 support on GeForce RTX 40. GeForce RTX 20 and GeForce RTX 30 support DLSS 2 and NVIDIA Reflex.

All games that support DLSS 3 are compatible with DLSS 2, but not vice versa, frame generation technology needs to be implemented. At the moment, the list of titles that already include support for this technology is very small, but we know that it will gradually expand, as it happened with the second generation of DLSS.

It should also be noted that in order to enable DLSS 3 and frame generation in games, these must use DirectX 12because this technology doesn’t work under DirectX 11. It’s not a problem because DirectX 11 is an outdated and obsolete API, but it’s important to be clear about it, because in games that allow us to choose between one and the other, we have to choose the first one to activate frame generation .