Overview
- NVIDIA showed Neural Texture Compression in a demo scene that dropped texture memory from about 6.5 GB with block compression to 970 MB with similar image quality.
- The system saves each texture as a compact latent and uses a small, deterministic neural network to reconstruct texels on demand inside the GPU.
- Decoding runs on dedicated matrix units such as Tensor Cores, XMX, or AI accelerators, which shifts work off the main shaders to protect base rendering performance.
- A related demo called Neural Materials encoded material behavior and reached up to 7.7x faster render times at 1080p in one scene without visible quality loss.
- The SDK is in public beta on GitHub and no shipping games support it yet, with DirectX’s Cooperative Vectors and rival demos pointing to cross‑vendor paths, and one unverified leak suggesting Sony may explore it for PlayStation 6 installs.