I've lost touch with the class of problems these machines deal with. It always used to be CFD, typically modelling things like nuclear explosions or dispersion of signals in complex materials like rocks for gas/oil exploration, radio/antenna signals stuff, aerofoil dynamics, fire front and climate.

Weather was about all of this combined with "reduce the cell size, increase the duration/lead time" but I (stupidly) believed we'd arrived at the floor: can't be radically more accurate, can't predict radically further into the future.

https://archive.fo/5hxV4 (for plugins that makes wsj grumpy).

https://www.wsj.com/tech/ai/chinese-supercomputer-overtakes-...

did it with only CPU's, no GPU's