VRAM cut: NTC demo
A Neural Texture Compression demo at GTC showed dramatic memory savings for graphics — up to a 7× VRAM reduction, with one example dropping from 6.5GB to about 970MB. (x.com). That kind of compression matters for running richer scenes on smaller GPUs and could change where and how developers deploy visual AI and game assets. (x.com)
At Nvidia’s recent GTC talk, a demo of Neural Texture Compression showed a single scene’s texture memory fall from about 6.5 gigabytes to roughly 970 megabytes while the image looked essentially the same. (videocardz.com) Textures are the image files that wrap a 3-D model — the bricks on a wall, the grain in a table — and they are the biggest consumer of GPU memory in modern games and visual apps. (research.nvidia.com) Neural Texture Compression, or NTC, replaces conventional block-based texture formats with a small neural decoder tied to each material. (research.nvidia.com) Instead of storing every texel at full resolution, the pipeline stores a compact neural representation — model weights and compressed feature data — and the GPU runs tiny neural networks to reconstruct texture samples on demand. (research.nvidia.com) Because NTC compresses all the maps that describe a material together (albedo, roughness, normal, etc.), it finds correlations the old block formats ignore and so delivers far higher apparent detail per byte. (github.com) In Nvidia’s Tuscan Villa demo, that approach let the company claim an almost sevenfold reduction in VRAM for textures while preserving more small-scale detail than ordinary BCn compression at the same memory budget. (videocardz.com) The trick is not lossless storage but efficient reconstruction: the neural decoder is cheap enough to run per-sample in the rendering loop, and the compressed representation is dense enough that the overall memory bill shrinks dramatically. (research.nvidia.com) NTC is not a single preset; Nvidia’s implementation exposes different inference modes that trade memory for runtime cost. (github.com) Independent testing and early reports show the most aggressive “on-sample” modes can reduce VRAM the most but may impose a performance cost — some testers report frame-rate hits on the order of tens of percent in that mode. (spilled.gg) Nvidia has released an NTC SDK on GitHub so developers can experiment with compressing materials and running the decoders inside a renderer; the project is labeled beta and already includes tools and sample renderers. (github.com) Because the decoder runs on the GPU, adoption depends on driver and API support and on whether game engines integrate the SDK; Nvidia has been updating LibNTC and working with DirectX cooperative-vector primitives to make inference efficient on popular platforms. (phoronix.com) For players and creators this is concrete: the same scene can fit into much less VRAM, which frees memory for higher resolution geometry, more objects, or ray-tracing buffers on modest cards, or it can let studios ship smaller downloads. (techpowerup.com 1) (techpowerup.com 2) The repository and Nvidia’s research notes include visual comparisons and quality metrics showing NTC delivering higher effective texel density than conventional GPU formats, and they provide the SDK and sample files for developers who want to try it now. (research.nvidia.com) If you want the code and samples, Nvidia’s RTXNTC project is published on GitHub under the Nvidia-RTX organization. (github.com)
