The comments left by our readers reflect something evident: Many don’t like the AI image reconstruction techniques used by the latest generation of graphics cards. However, whether you like it or not, this is the way to go, especially if you want to play at high resolutions and with the highest possible image quality.
One of the main contributions of the GeForce RTX 50 recently presented by Nvidia is the Deep Learning Super Sampling (DLSS) technology. Although this revision of the image reconstruction strategy using AI systems is only fully available for the GeForce RTX 50, it also includes innovations compatible with the RTX 40, 30, and 20. According to Nvidia, it multiplies performance by up to 8 times using native resolution rendering as a reference.
The Heart of DLSS 4: Multi Frame Generation
Nvidia introduced DLSS technology with the first generation of GeForce RTX graphics cards, promising users to enjoy their games at a higher frame rate, even if their graphics needs are very demanding—even with ray tracing enabled. This innovation relieves the GPU of some of the work in rendering images to increase the frame rate without compromising graphics quality.
The idea is ambitious, and as users can guess, the technology that makes it possible is complex. Nvidia’s image reconstruction technique relies on real-time analysis of game frames using deep learning algorithms. The strategy Nvidia uses to reduce the load on the GPU involves rendering at a lower resolution than the output resolution ultimately delivered to the monitor.
With DLSS, the rendering resolution is lower than the final output resolution delivered by the graphics card to your monitor.
This reduces the load on the graphics processor, but it also requires a procedure for scaling each frame from the rendering resolution to the final resolution. Moreover, it must be done efficiently. Otherwise, the effort avoided in the previous stage could appear in this phase of image generation.
Here is where the AI systems developed by Nvidia and the GPU’s tensor cores come into play. The graphics engine renders the images at a lower resolution than expected, and then DLSS technology scales each frame to the final resolution using a deep learning sampling technique to recover as much detail as possible.
|
|
geforce rtx 50 |
geforce rtx 40 |
geforce rtx 30 |
geforce rtx 20 |
|---|---|---|---|---|
|
Multi frame generation |
Yes |
No |
No |
No |
|
Enhanced frame generation |
Yes |
Yes |
No |
No |
|
Improved ray reconstruction |
Yes |
Yes |
Yes |
Yes |
|
upgraded super resolution (beta) |
Yes |
Yes |
Yes |
Yes |
|
Advanced deep learning antialiasing (beta) |
Yes |
Yes |
Yes |
Yes |
The table lists the leading technologies that make up DLSS 4. As expected, all of them are available in the new GeForce RTX 50, but only some were available in previous generations of Nvidia graphics cards. The latest revisions of the super-resolution and deep learning antialiasing technologies are still in beta development. However, the final version of these two innovations will likely be available soon. Antialiasing is the process used to reduce the jagged edges of objects in each frame.
According to Nvidia, Multi Frame Generation technology is primarily responsible for the GeForce RTX 50’s 8x performance boost over native resolution rendering. It uses AI features to generate up to three frames per rendered frame to achieve this. This feature is only available on the RTX 50 because it requires specific functional hardware units Nvidia has introduced in its new graphics cards.
The new frame generation model using AI tools is 40% faster than the one used in DLSS 3. It uses 30% less VRAM memory and only needs to be executed once per rendered frame to generate multiple frames. Nvidia states that its engineers have managed to speed up frame generation by replacing some of the hardware present in the GeForce RTX 40 with a more efficient AI model. Interestingly, the combined work of the latter model and the primary model of the Multi Frame Generation technology significantly reduces the computational cost of generating additional frames.
The following graph reflects the performance of the GeForce RTX 5090 at 2160 p with several relatively recent games (although not all of them) and in three different scenarios: without DLSS, with DLSS 3, and with DLSS 4. As you can see, this last image reconstruction technology delivers the highest performance in all titles. In some of them, such as Alan Wake 2 or Black Myth: Wukong, it manages to multiply the rendering performance at native resolution by 8.2 times without DLSS.
The image below describes a key feature of DLSS 4 that GeForce RTX 50 and RTX 40, 30, and 20 take advantage of. Previously, DLSS technology used convolutional neural networks to generate new pixels by analyzing adjacent pixels and monitoring changes in each region in successive frames. However, DLSS 4 implements a new transformation model that roughly evaluates the relative importance of each pixel within each frame and over several consecutive frames.
The main advantage of this new model is that it can “understand” each frame much more accurately, resulting in higher quality pixels, sharper images, more detail during camera movements, and smoother edges on objects on the screen. Ray tracing, super-resolution, and antialiasing technologies exploit this new transformation model. Hence, as mentioned above, the GeForce RTX 20 and all subsequent series benefit from it.
Images | Nvidia
Related | Nvidia’s Project Digits is the Start of Something Big: A Desktop Supercomputer for AI Systems
View 0 comments