AMD looking to improve texture compression with AI-powered algorithm

While the mainstream is trying to figure out the ethics of large language model usage, scientists and even chip makers are putting AI to good use and yielding palpable progress across numerous fields. Neural networks in particular are very useful in this sense. Nvidia is already using neural network tech in its GPUs with DLSS plus ray reconstruction leading to clear benefits for the end users. Last year, Team Green demonstrated how neural networks could also be employed to improve texture compression, but no further developments that could benefit gamers have been announced in the meantime. AMD has revealed that it is also working on a similar technology and this will be detailed next week at the 35th Eurographics Symposium on Rendering in London.

It is not yet clear how AMD’s tech differs from the Nvidia one. Both seek to improve upon the existing Block Compression method, but Nvidia’s version seems to take a completely different approach, whereas AMD is using Block Compression as the basis. Furthermore, Nvidia’s method is focusing on increasing the texture resolution by four times while maintaining the same data size from Block Compression, when AMD’s goal seems to be data size and load times reduction.

AMD’s announcement tweet additionally mentions that such a technique would require “unchanged runtime execution” allowing for “easy game integration.” This seems to suggest that AMD’s compression method is easier to adopt for game developers, yet we still need to see how all this would come about. Direct driver integration could greatly simplify things, but this would restrict the new compression to AMD’s cards, most likely only for newer generation GPUs. Alternatively, AMD could open-source the code, allowing integration for its own GPUs as well as for Nvidia and Intel cards.