AMD's Neural Block Compression Tech Is Coming To Save Gamers From Huge Downloads
Given that the presentation hasn't happened yet, we don't know that much about the technique, but based on this diagram in the tweet, we can infer a bit. It appears that AMD's idea is to compress the many different texture layers used on surfaces in modern games (like UV maps, specular maps, normal maps, roughness maps, and so on) into a single layer that is then read out by AI back into the various layers required by the game engine.
If that all sounds a bit familiar, it's probably because NVIDIA already announced a neural texture compression method more than a year ago. However, NVIDIA's method seems to be a bit different, as it requires game engine integration. AMD says that its method offers "unchanged runtime execution," meaning that developers don't have to modify their game code to use it. We're not quite clear on how that works out, but it certainly sounds promising for adoption. NVIDIA's method has yet to be used in any game.
We're interested to see exactly how AMD's method differs from NVIDIA's method, so look for more coverage of this tech when AMD unveils it next week.