Cerrebrus solved the power problem for wafer scale chips. So now for a mere 2 million USD you can get a exaflop neural net system that sits on a desk top and chanels 20kwh of heat- like 10 hair dryers running full blast. They focused on neural net tech specifically. They are on their 2nd gen of commercial systems. There are apparently no competitive commercial hardware systems. Google's version is supposedly half the flops and takes up a tennis court and uses a massive amount of energy. Google is pairing quantum computers though. Their goal for a decade has been quantum AI neural net machines in the cloud replacing all other computers essentially- but they recently lost or moved the guy responsible for that strategy amd they lost the architect of their FSD Waymo tech to Tesla and Tesla started a more advanced general approach. Google's quantum AI neural net pairs CPU, Neural Net and a Quantum processor- GPU is surely in there too to some extent and DSP etc., or maybe not. All that is missing from this kitchen sink is the fabled time processor to solve halting problem issues. Its interesting how all this stuff is compressing or integrating onto single dies in an acceleration of SOCs even including the DRAM. Apple's M1 includes all (minus a quantum component) with the actual system DRAM on a single chip. Microsoft is trying to put a general purpose quantum computer on a chip. With wafer scale maybe everything including SSD can go on a single chip? Wonder if the entire system board will reduce to a wafer chip with a cartridge port on a side. What a wonderful kind of milspec tough packaging. We could see that in a few years.
