Nvidia recently issued a press release, showing a new texture compression algorithm Neural Texture Compression (NTC). Using this algorithm compared to the block compression algorithm, one can improve video resolution by 4 times and reduce video memory consumption by 30%.
Block Compression (BC) is a technique used to reduce the amount of video memory required to store color data. By storing some colors at their original size and using other encoded patterns, you can dramatically reduce the video memory required to store images. Using compressed textures does not cause performance degradation because the hardware automatically decodes the compressed data.
NTC uses tensors (three dimensions) to represent textures, and the new method aims to compress multiple channels and mipmaps (textures of different sizes) together to achieve better quality and bit rates than JPEG XL or AVIF. Unlike traditional block compression algorithms that require special blocks, the NTC algorithm uses matrix multiplication techniques performed by modern GPUs themselves.
Neural textures can be rendered in real-time, with 16x more texels than the BC method and a 4K rendering speed of 1.15 ms (measured on RTX 4090). More information on this research will be presented at SIGGRAPH 2023 on August 6.
