AMD introduced the NTBC AI texture compression method: details and prospects

AMD has announced a new AI-powered texture compression method called NTBC (Neural Texture Block Compression). The information was presented at the 35th Eurographics Rendering Symposium and details have now been published on the website GPU Open.

NTBC's main goal is to reduce the size of game files, which have grown significantly in recent years and can reach up to 150 GB. The main reason for the increase in file sizes is high-quality textures for 4K displays and other high-resolution devices.

NTBC uses a neural network to compress textures, reducing their volume by up to 70% while maintaining acceptable quality and without changing shaders. Researchers Shin Fujieda and Takahiro Harada from AMD explained that NTBC uses multi-layer perceptrons (MLPs) to encode all texture data into a single material, allowing it to achieve lower bitrates compared to standard BC.

The NTBC method does not require changes to shaders because the network weights are stored on disk and loaded into memory, followed by decompression of the texture data, which is then copied to video memory (VRAM). This makes the method easy to integrate into existing graphics pipelines. Multi-level feature grids and quantization-aware learning are used to optimize the model and reduce storage costs.

While NTBC promises a significant reduction in texture volume, its use may result in blurred detail or blocking artifacts due to the use of lower resolution meshes. Possible solutions include various encoding techniques such as texture focusing or local positional encoding. AMD researchers are also considering extending NTBC to BC6H and BC7 formats.

An important aspect of NTBC is the moderate computational load when loading textures. Decompression execution time is estimated to range from 27.31 to 49.84 ms depending on the choice between a conservative (using two models for RGB and single-channel textures) and an aggressive approach.

Thus, NTBC represents a promising method for reducing game file sizes, which is especially important in the face of ever-increasing data requirements.