NVIDIA Neural Texture Compression Slashes VRAM Usage Over 80% In Games
In the current market environment, efficient memory usage has become increasingly more important. Even with the scene compressed down to 670 MB, NVIDIA's NTC still shows a remarkable bump in fidelity compared to other texture compression techniques. The dinner table scene pictured below uses traditional rendering techniques (downscaled BCn textures) in the top half and NTC in the bottom half. As you can see, the lower portion of the image has much more detail and looks more photorealistic. And that's before taking advantage of other techniques like DLSS.
Besides the improved texture compression, NTC also offers support for multi-layered neural materials. Like textures, physically-accurate materials are already commonly used to simulate subtle light diffusion, reflections, and other ways light may interact with a surface. Full neural rendering of textures and materials also sees massive performance and fidelity gains. While "inference" is indeed being used with these neural techniques, no "generative" AI is in use until you step up to DLSS 5.
So, part of what makes DLSS 5 possible isn't simply slapping a generative AI post-process frame on top of existing data. Beneath the surface, these neural rendering techniques are also needed to reduce memory usage, which makes more memory available for other things. It's up to developers to figure out how they want to utilize and implement NTC. Especially by itself, the performance gain offered by NTC (nearly 70% less render time for the same degree of fidelity) is impressive.
For an in-depth look at neural rendering for textures and materials, the original NVIDIA keynote is embedded below.



