Last week, Brian Bailey published an interview with Professor Madhavan Swaminathan who is the Director of the Interconnect and Packaging Center (IPC) at Georgia Tech in Atlanta. The topic of the interview was cooling of 3D IC devices. It’s no secret that cooling will become an issue when you stack a bunch of hot die together and bond them. The heat has to go somewhere and the heat from the middle die has nowhere to go but through the outer die in the stack. For a lot of designs, that’s not going to work. The stack will overheat.
What’s the solution? Di-Hydrogen Oxide, aka water.
Water cooling is both the answer and the non-answer.
As Bailey asks: “So you are saying water cooling is the only solution but at the same time it is not viable.”
Swaminathan’s reply: “Exactly. At this point in time, people have demonstrated, including Georgia Tech, that you can cool the chips by creating micro-channels, creating hollow TSVs that serve as a pipe to pump water from the printed circuit board into these micro channels, and you get that heat out and you basically build up a closed loop system. And using that, you can bring the temperature of the ICs down, but obviously from a system standpoint, this may not be a cost effective solution, and people are looking at other ways to solve the problem, and to this day, I haven’t seen a rival solution yet.”
And that’s sort of where we are today with 3D IC cooling. It’s a work in progress.
However, that’s not the end of today’s blog post because yesterday, I published a blog entry in the Denali Memory Report about a 3-petaFLOP supercomputer called the SuperMUC that runs on 3.5 MW of electricity. As you might guess, there are some cooling challenges with the SuperMUC’s 18,432 Intel Xeon Sandy Bridge-EP multicore microprocessors and 324Tbytes of Samsung Green 30nm-class DDR3 SDRAM.
IBM engineers have developed a water-based cooling solution for this class of machine. It’s called Aquasar, which reportedly cuts the energy needed for cooling by 40% by using 60°C (or perhaps 45°C depending on the reference) water for cooling.
For the purposes of this blog post, the important point about Aquasar cooling is that the working fluid flows through microchannels cut into copper blocks that are directly mounted to the processor die. All 18,432 of them. Here’s a photo of an Aquasar processor module:
What’s this got to do with 3D IC assembly? Well, this article explains the Aquasar cooling system in detail and makes a point of discussing the ability to cut cooling channels where they will do the most good. In other words, if there’s a hot spot on the processor chip, you put more cooling fluid in that region to pull out more heat. The same idea will work with cooling channels cut directly into the silicon die in a 3D IC stack—in concept.
IBM’s Aquasar cooling technology is not a practical production technology for most of today’s systems. You can use lots of expensive system-level ideas in supercomputer design that cannot be applied in the designs intended for everyday life. However, Aquasar is a walking, breathing proof of concept. You certainly can pipe coolant through microchannels to cool semiconductors. It’s no longer a question of if, merely when.