NVIDIA's New DLSS Transformer Reduces VRAM Usage by 20%

NVIDIA recently rolled out the DLSS Transformer model as part of its 310.3.0 SDK. Unlike previous DLSS versions that relied on convolutional neural networks to infer missing pixels, the new Transformer-based approach evaluates relationships among all pixels across multiple frames. This delivers more consistent upscaling while dramatically reducing memory usage.

NVIDIA’s engineers claim the new model trims VRAM requirements by approximately 20%. In one example cited, running DLSS at 1080p now consumes about 87.8 MB of VRAM compared to 106.9 MB in the previous SDK. Similar reductions apply at 1440p, 4K and even 8K resolutions. For gamers running GPUs with 8 GB or less of memory, this decrease in resource demands translates into smoother performance and the ability to enable additional graphics features without sacrificing frame rate.

Game developers and engine partners are expected to integrate the DLSS Transformer model into their titles over the coming months. Early tests report crisper edges, more stable frame rates and consistently high upscaling performance.

*Source: TechPowerUp*