DLSS 4's new enhanced Super Resolution now uses a lot less VRAM

Last week, NVIDIA's new DLSS 4 Super Resolution, which utilizes a more advanced and complex Transformer model, exited beta. Debuting alongside the GeForce RTX 50 Series earlier this year (with support for GPUs from the RTX 20, 30, and 40 Series cards), the new DLSS 4 Super Resolution sets a new standard for AI-powered upscaling that delivers incredible image quality that is often sharper and more detailed than native rendering.

2

It's become a key selling point for the GeForce RTX 50 Series thanks to it already being available in over 125 games and apps. With it officially exiting beta and becoming part of the DLSS Super Resolution and Ray Reconstruction SDK 310.3.0, it will start becoming available in more and more titles, with the older, existing CNN (Convolutional Neural Network) variant set to be phased out.

With an updated version of the DLSS SDK for game developers now available, which includes the official release of DLSS 4's new Transformer model for Super Resolution, we've also learned that NVIDIA was able to reduce the VRAM requirement for running the latest model.

First spotted by Videocardz, the transition from beta to full release was more than a formality, as NVIDIA has optimized the new AI model to use around 20% less VRAM. It's still more intensive than the existing CNN model, but instead of using nearly double the VRAM, that figure has now dropped to around 40% more.

As seen in the chart below, we also learn that moving from 1080p to 1440p, 4K, and beyond, dramatically increases the VRAM footprint of DLSS 4 Super Resolution. Going from 143.54 MB at 1440p to 307.37 MB at 4K is a 3.5X increase.

Although a few megabytes at lower resolutions probably won't translate to an additional performance that you can see, this is still fantastic news alongside the new DLSS 4 Frame Generation model, which uses 30% less VRAM, and AI-powered texture compression on the horizon, it's great to see solutions that go beyond 'throw more VRAM at it."