DLSS, the latest incarnation of the transformer model consumes less vram

Not only did he come from the beta, but also the latest version of the Transformation model used for the DLSS (Deep Learning Super Sampling) It introduces optimisations aimed at managing video memory. What does it mean? That Vs name form. With the latest version of theSDKidentified as 310.3.0The VRAM needed for scaling up through the transformation model Reduced by around 20% compared to the previous version.

This improvement was reported by Nvidia itself in the official documentation for developers. It is An important step forward for users with graphic maps equipped with 8 GB Vram or not.

With the update in question, the transformer model requires 85.77 MB VRAM on 1080p, against 106.9 MB of the previous version. For a comparison, the old CNN model (Convolutional Neural Networks) was 60.83 MB at the same resolution, but the return of a lower image quality.

In 4kVram’s consumption of the new transformer comes to 307.37 MB, about 80 MB less Compared to the previous implementation. Although they are relatively modest figures with regard to the VRAM of the GPUs, these savings become more relevant for ultra -high resolutions, such as 8K, where the updated model consumes slightly more than 1.2 GB.

It is important to underline that these optimisations apply exclusively to the upscaling module. The generation of the frames (frame gene), another central component of the DLSS, has already benefited from a memory saving of 30% with the transition to version 4, for example, Nvidia has reported that in Warhammer 40,000: Darktide, the new generation of frame about 400 MB Vram less than 4K than the previous consumes.

Even if memory reductions can seem marginal, every detail counts, especially for those who use Mid -Range GPU with limited sources – and we are not just talking about the last generation.