H100s also has custom async WGMMA instructions among other things. From what I understand, at least the async instructions formalize the notion of pipelining, which engineers were already implicitly using because to optimize memory accesses you're effectively trying to overlap them in that kind of optimal parallel manner.