An interesting post on RTF (https://www.realworldtech.com/forum/?threadid=205830&curpostid=205830) is discussing how Apple might be manufacturing the Ultra. Now, I have very little clue about chip technology, but from what I gather, a core challenge for modern multi-chip technologies is packaging multiple chips together. What is commonly used is an interposer - an additional, larger, simpler chip that acts like a sort of a communication network and on top of which the actual chips are mounted. This works, but is expensive and not very space-efficient. The post describes various state of the art technologies in great details and speculates that Apple might be using a different approach that has its roots in packaging of mobile chips. If I understand it correctly, it is argued that they are using some smart tricks to precisely position chips on a carrier wafer and lock them together, which essentially creates a new “fake wafer” with chip innards exposed. They can then use more or less conventional technique to put an additional interconnect level over this “fake wafer” - just as it is normally done with any chips (this is called BEOL). The result is a tightly integrated, compact structure that has much better signal efficiency and electrical properties than interposer technology, while also being much cheaper as you don’t need to manufacture the extra interposer chip. This is literally a “monster chip” built from multiple chips, bypassing yield issues and other problems.
If this speculation is correct, then it spells out very good news for Apple Silicon. This is a very impressive technology that is inherently very scalable. In the future Apple could probably use it more aggressively to build up very flexible configurations from individual dies. The thing is, the chips connected this way don’t need to be identical. Imagine smaller building blocks: CPU/GPU clusters, cache blocks etc., using smaller dies with high yields, being packaged together on demand and glued into one big SoC. One could build some truly monstrous systems without incurring the high cost and suffering from low yields usually associated with them.
P.S. the author of the RTW Post is probably Maynard Handley who previously published a lot of in-depth information on Apple Silicon, most notably this insanely cool investigation. He is a member here on MR with the handle name99.
If this speculation is correct, then it spells out very good news for Apple Silicon. This is a very impressive technology that is inherently very scalable. In the future Apple could probably use it more aggressively to build up very flexible configurations from individual dies. The thing is, the chips connected this way don’t need to be identical. Imagine smaller building blocks: CPU/GPU clusters, cache blocks etc., using smaller dies with high yields, being packaged together on demand and glued into one big SoC. One could build some truly monstrous systems without incurring the high cost and suffering from low yields usually associated with them.
P.S. the author of the RTW Post is probably Maynard Handley who previously published a lot of in-depth information on Apple Silicon, most notably this insanely cool investigation. He is a member here on MR with the handle name99.
