Quote:AMD Neural Texture Compression

Next week, AMD is set to showcase its method of using neural networks for texture compression. 



It should be reminded that last year, NVIDIA presented a very similar technology in one of their research papers. The company proposed “Random-Access Neural Compression of Material Textures,” which, according to the paper, would allow neural algorithms to increase resolution by four times compared to Block Compression methods.

AMD appears to be working on technology to compress textures using neural networks as well. A presentation titled “Neural Texture Block Compression,” scheduled for July 2nd and delivered by S. Fujieda and T. Harada from AMD, is said to focus on new compression methods that would allow data to be reduced. As the title suggests, it would use Block Compression.

AMD promises that this technology will be easy to implement because it uses unchanged runtime execution. However, it remains unclear whether this technology can be implemented at the driver level or if developers will need to add the necessary code. Additionally, the required hardware for this technology is still unknown.

NVIDIA’s similar technology has not yet been adopted by any game. While AMD’s claim of easy game integration is promising, it does not guarantee widespread use. Open-sourcing the code or incorporating the technology into the FidelityFX software stack would likely make it more attractive to game developers.

Neural networks are expected to play an even greater role in gaming in the future. Recently, NVIDIA’s CEO hinted that the company might eventually focus on improving the quality of textures and using AI to generate higher-resolution objects in games.
...