Obviously not much specific detail at this stage, but it strikes me as odd that there's no points from Cerebras that seem to address the obvious concern with fabrication errors - I'd always understood that by having many small chips that mitigated costs, because one single chip could be toast but you'd still have a substantial number of unaffected ones to use.
Unless they've got some special solution up their sleeves, it seems like one error would knock out the whole chip.
Also, by needing it to be square, won't they fail to utilise the sections between the edge of the square and the edge of the wafer's circular edge?
The whole thing is probably then immersed into some kind of non-conducting fluid (mineral oil or fluorinert would be my two first guesses), probably vertically to take advantage of thermal circulation. Mechanical circulation of the coolant would likely be very gentle so-as not to disturb the wire bonds, unless they have something special to keep them in place (maybe the bonds and edge of the chip are encapsulated leaving the center bare?).
All of this is just guesses from somebody not an expert or even ever touched stuff like this, just based on what little I know about how normal sized chips are packaged.
1.2 Trillion on 8.5" squared gives 26M/mm2. Impressive. But would take cryo lab to cool, i think.
