NVIDIA’s Upcoming DLSS “Transformer Model” Will Slash VRAM Usage by 20%, Bringing Smoother Performance on Mid-Range GPUs
NVIDIA’s next big update with DLSS would likely be the switch towards a “Transformer Model”, which is expected to bring several improvements, including lesser VRAM utilization.
NVIDIA DLSS Transformer Model Would Come With Less VRAM Usage, Helping Gamers With 8GB or Lower GPUs
NVIDIA previously utilized CNN (Convolutional Neural Networks) for its DLSS technology, but it seems like Team Green has now switched to a more advanced transformer model approach, which has brought in a wide range of benefits, from improved image quality to superior ray reconstruction. The DLSS Transformer model came out of beta a few days ago, and it is claimed that NVIDIA has brought in VRAM optimizations with the new model as well, with the changes mentioned extensively in the DLSS Programming Guide (via Videocardz).
Well, this marks an average of around 20% lesser VRAM requirements with the newer DLSS Transformer model in the 310.3.0 SDK, and this difference is consistent across all resolutions, showing that the update would have a wider impact. This is certainly a great addition, especially for gamers looking to capitalize on the power of DLSS with confined VRAM capacities such as 8GB or lower. Considering how NVIDIA’s recent moves haven’t done much for budget gamers, the DLSS improvements are definitely something to appreciate.
For a quick recap on what the Transformer model brings to the market, it uses a vision transformer that evaluates all the pixels in a particular frame and understands the importance of individual pixels. This is repeated across multiple frames to generate detailed pixels for enhanced visuals, and it is known to increase the parameters by 2x, and rendering compute by 4x, showing that DLSS Transformer model would bring noticeable improvements to the upscaling experience.
Given that the DLSS Transformer model is out of beta, we should expect an official update deploying the technology within the upcoming months.
