NVIDIA’s GDC talk shows neural rendering is not just DLSS 5, it can also cut VRAM use by 85%
© NVIDIA Shortly after Jensen presented DLSS5 at GTC 2026, NVIDIA also discussed neural rendering in practice during a GTC 2026 talk just few days later. The talk focuses on more technical level and how this technology not only changes rendering pipelines, but also how it can be used in something else like gaming.
DLSS 5 is only one part of neural rendering, and it sits on the side where machine learning is applied to the final rendered result. The GDC session instead focused on using small neural networks inside the rendering pipeline itself to decode textures, evaluate materials, and reduce memory traffic. Think, smaller neural engines dedicated to special tasks instead of one final filter.
Source: NVIDIA
The easiest example was Neural Texture Compression (NTC) NVIDIA showed its Tuscan Wheels scene dropping from about 6.5GB of VRAM with traditional BCN-compressed textures to 970MB with NTC, while keeping image quality close to the original. The company also said that at the same 970MB memory budget, NTC preserved more detail than regular block compression. That kind of optimization means smaller installs, smaller patches, less download bandwidth, and more room for detailed assets on the same GPU. This may ring a bell to you, because we already discussed NTC on VideoCardz a while back .
Source: NVIDIA
Another interesting AI-rendering angle is Neural Materials , which follows the same logic. Instead of carrying a large stack of separate texture data and running heavier BRDF math, NVIDIA compresses material behavior into a compact latent representation and decodes it with a small network. In the demo, a material setup with 19 channels was reduced to eight channels, and NVIDIA reported 1.4x to 7.7x faster 1080p render times in that test scene. That is less about inventing new visuals and more about storing and evaluating the same material data in a lighter way.
Source: NVIDIA
The DLSS 5 drama has been about fears that AI could push games toward a generic look, interfere with art direction, or change the final image in ways players do not want. Jensen Huang has already responded to that criticism after DLSS 5 was labeled “AI slop” by some users and developers.
NVIDIA might have made a mistake by showing DLSS5 this early, and instead of focusing on benefits for gamers, such as lower VRAM use, higher quality textures, and small updates to game rendering pipeline, they decided to promote a technology that may change the game entirely.
I think I prefer the discussion on how AI can help games achieve better optimization, quality, or even performance, rather than AI that changes the entire look of the game, for better or worse.
Source:
