Diamonds, lasers, and oil aren’t the first things you may think of when considering ways to keep chips and computers cool. But as modern chip designs pack and stack more transistors into ever smaller spaces, heat has emerged as a critical problem.
To solve it, the semiconductor industry is throwing everything at the wall. What sticks could enable the scaling of not only AI data centers but also a host of applications in consumer electronics, communications, and military equipment.
As Senior Editor Samuel K. Moore explained to me between bites of a cold tongue sandwich at the 2nd Ave Deli, near IEEE Spectrum’s office, better thermal management is essential for next-generation nodes.
“As we start doing more 3D chips, the heat problem gets much worse,” said Moore, who has been covering semiconductors on and off for a quarter century.
For the special report in this issue, Moore teamed up with Associate Editor Dina Genkina, who oversees our computing coverage. They talked to engineers at IEEE conferences like IEDM and Supercomputing about how technologists are getting the heat out in new and surprising ways.
“As we start doing more 3D chips, the heat problem gets much worse.” —Samuel K. Moore
The first step to solving an engineering problem is characterizing it precisely. In “Will Heat Cause a Moore’s Law Meltdown?”, James Myers, of Imec in Cambridge, England, describes how transistors entering commercial production in the 2030s will have a power density that raises temperatures by 9 °C. In data centers where hot chips are crammed together by the millions, this increase could force hardware to shut down or risk permanent damage.
In “Next-Gen AI Needs Liquid Cooling”, Genkina takes readers on a deep dive into four contenders to beat this heat with liquids: cold plates with a circulating water-glycol mixture attached directly to the hottest chips; a version of that tech in which a specialized dielectric fluid boils into vapor; dunking entire servers in tanks filled with dielectric oil; and doing the same in tanks of boiling dielectric fluid.
Although liquid cooling works well, “it’s also more expensive and introduces additional points of failure,” Moore cautioned. “But when you’re consuming kilowatts and kilowatts in such a small space, you do what you have to do.”
As mind-blowing as servers in boiling oil may seem, the two other articles in this issue focus on even more radical cooling technologies. One involves using lasers to cool chips. The technique, outlined by Jacob Balma and Alejandro Rodriguez from the Minnesota-based startup Maxwell Labs, involves converting phonons (vibrations in a crystal lattice that carry heat) into photons that can be piped away. The authors contend that their technique “can target hot spots as they form, with laser precision.”
Meanwhile, Stanford’s Srabanti Chowdhury takes a blanket approach to the heat problem, swaddling transistors in a polycrystalline diamond film. Her team’s technology has progressed remarkably fast, reducing diamond-film growth temperatures from 1,000 °C to less than 400 °C, making it compatible with standard CMOS manufacturing.
None of these solutions comes cheap, and so the future of chips is going to be expensive as well as hot. That probably doesn’t faze the big AI companies sitting on giant piles of investors’ cash. As Moore pointed out as he polished off a pickle, “AI’s demand for chips is sort of unlimited, so you’ve got to do things that you wouldn’t have thought of doing before and swallow the expense.”
From Your Site Articles
Related Articles Around the Web