Alongside their recently unveiled RTX 50 series GPUs, Nvidia also showcased the upcoming DLSS 4 AI frame generation technology, which promises significant improvements while ensuring considerably lower computational costs. Naturally, we assumed that the RTX 50 series would have the cores to handle the demanding needs of these upcoming AI advancements, but it looks like all RTX cards, even the RTX 20 series, will benefit from these advancements in one way or another.

According to Nvidia’s blog post, DLSS 4 introduces a major architectural change that will allow all RTX GPUs to deliver significantly improved image quality for Super Resolution, Ray Reconstruction, and DLAA. This change comes in the form of the first real-time use of a transformer-based model in the graphics industry, capable of enhancing image quality, lighting details, and stability in supported titles by a substantial margin.

New Transformer Model in DLSS 4 uses 2x the parameters

Previously, DLSS used Convolutional Neural Networks (CNNs), which could only focus on small portions of an image at a time, analyzing local regions of pixels to generate better details. It worked by predicting what missing or low-quality pixels should look like based on its training. However, after six years of development, the CNN model reached its limits in performance and quality.

Now, DLSS 4 replaces the CNN with a transformer-based model that is capable of considering the entire image at once, analyzing all pixels together. This makes the AI better at handling details and improvements across the entire frame. Additionally, the transformer model uses double the number of parameters compared to the old CNN model, giving it a better understanding of a scene and helping it make quick decisions.

As a result, it can generate more stable images with less ghosting and better detail, especially during movement or changes in the scene. It also makes motion appear smoother and sharper, without the distortions that older upscaling methods sometimes caused. Confused? Check out the Alan Wake 2 demonstration below to see how the new transformer model for Ray Reconstruction improves stability on the highly detailed chainlink fence and reduces ghosting on the fan blades.

Similarly, the transformer model for Super Resolution shows promising results, offering better temporal stability, less ghosting, and higher detail in motion. Nvidia mentions this will be released as a beta to allow users to explore the improvements and provide feedback before the official release.

All RTX GPUs will be able to take advantage of this new Transformer Model

Interestingly, this new Transformer Model will not be limited to RTX 50 series GPUs but will also be compatible with the RTX 40 series, RTX 30 series, and even RTX 20 series GPUs. All four series will be able to leverage the following technologies (all of which have been enhanced with DLSS 4):

• DLSS Ray Reconstruction
• DLSS Super Resolution
• DLAA

However, there are a few exclusive features on the list as well. This includes the enhanced DLSS frame generation, which will offer faster frame rates while utilizing less VRAM, and the all-new Multi Frame Generation feature. The enhanced DLSS frame generation feature will only be compatible with the RTX 40 series and RTX 50 series GPUs, whereas the Multi-Frame Generation feature is exclusive to the RTX 50 series lineup. This means that if you own an RTX 40 series GPU, the only thing you’re missing out on is the Multi-Frame Generation, while all other enhancements are available to use without needing an upgrade.

That said, Nvidia also states that GeForce RTX users will only be able to experience the new transformer-based DLSS Ray Reconstruction, DLSS Super Resolution, and DLAA in supported games and apps. We are still yet to see the level of improvements these enhanced DLSS techniques will offer on older cards like the RTX 3070 or even the RTX 2070.

About the Author

Hassam Nasir

Hassam boasts over seven years of professional experience as a dedicated PC hardware reviewer and writer.