DLSS Transformer from NVIDIA Reduces VRAM Usage by 20%

As part of the 310.3.0 SDK, NVIDIA has formally released their DLSS Transformer model, which combines significant video memory reduction with sophisticated upscaling. By reducing VRAM utilization by 20%, this update directly targets the demands of players using graphics cards with 8 GB or less. The Transformer method assesses the associations between every pixel in a frame and applies that knowledge over several frames, in contrast to earlier DLSS versions that used convolutional neural networks to predict missing pixels. Even though this technique has become more complex, NVIDIA’s engineers have improved its memory management practices to keep resource requirements low and produce images that are crisper and more reliable.

The improvements are noticeable when tested at standard resolutions: DLSS now uses just 87.8 MB of VRAM at 1080p, compared to 106.9 MB in the previous SDK release. Similar 20% reductions are seen at 1440p, 4K, and even 8K. These savings result in better performance and the ability to enable richer graphics features without sacrificing image quality for users with GPUs that have limited memory. Over the next few months, game developers and engine partners may anticipate incorporating the DLSS Transformer model into their titles and tools. Preliminary testing has already shown improved upscaling performance, more constant frame rates, and clearer edges.