Nvidias Updated DLSS Reduces VRAM Usage by 20% with Enhanced Upscaling Efficiency

Nvidia has streamlined VRAM usage in its latest DLSS SDK version. The DLSS 310.3.0 update facilitates a 20% reduction in video memory consumption during image upscaling. This improvement is attributed to optimizations made to the Transformer model, as reported by Tom’s Hardware.

Previously, the DLSS Transformer model used nearly twice the amount of memory compared to the Convolutional Neural Network (CNN) it replaced. Now, that gap has been narrowed to just 40%.

In the 310.3.1.0 SDK version, the VRAM requirement for Transformer models at 1080p is 87.77 MB, down from 106.9 MB in the previous version. It’s worth noting that DLSS memory consumption scales linearly with resolution: higher resolutions demand more video memory. For instance, when utilizing DLSS at 4K, the Transformer model requires 307.37 MB, which is 3.5 times the amount needed at 1080p.

This memory optimization specifically pertains to image upscaling, not frame generation. However, DLSS 4 introduced enhancements in frame generation that have led to a 30% reduction in VRAM usage. Frame generation typically requires significantly more VRAM compared to the upscaling aspect of DLSS. For example, in a blog post, Nvidia stated that Warhammer 40,000: Darktide consumes 400 MB less VRAM at 4K resolution when using DLSS 4 for frame generation compared to DLSS 3.