The M2 Ultra chip is unusual not because it's a single chip made of two halves, but because the halves are centrosymmetric images of each other instead of translation. Assuming the wafer is exposed first for the odd-numbered patterns, flipped 180 degrees, then for the even-numbered ones,
- Can you even do that without a custom machine that does the flipping in the lithography lines?
- Can you shut down the CO2/ArF laser for 50% of the time with ~1s each time to not expose the opposing patterns during a pass of a wafer?
- Assuming the newest generation of lithography machines perform as advertised and achieve ~1 nm superimposing accuracy, does that mean it's possible to have transistors instead of just micron-scale wires riding the boundary of 2 different exposures? Theoretically, does this allow exposing the whole wafer as a single monolithic chip with different masks for different regions?
EDIT: After some research, it appears I'm misled by a screenshot of one of Apple's patents here
into thinking that it's something other than 2.5D packaging. Still, my original questions stand.
EDIT2: Let me clarify what I meant by the original question. Suppose I have a wafer with 100 patterns to expose. Is it possible to first expose 1,3,5,...,99, get the wafer out of the lithographer, use a mechanical arm to rotate the whole wafer 180 degrees, send the wafer back into the lithographer, then expose 2,4,6,...,100, such that the adjacent patterns, e.g. 5 and 6, are centrosymmetric instead of a simple translation of each other. Note that this would NOT need any kind of interposer, change of masks, and the wires could be in say M1 or M2, which means they could be nm in pitch instead of um.
Please note that if you want to make a chip like M2 Ultra, the two halves have to be centrosymmetric.