Intel-led Team Demonstrates First Chip-Scale Thermoelectric Refrigerator

An integrated thermoelectric device cools a hot spot on a much larger chip

3 min read

28 January 2009—Researchers at Intel, Arizona State University in Tempe, RTI International, and Nextreme Thermal Solutions reported Sunday that a small thermoelectric device embedded in a chip package could cool a much larger chip. The thermoelectric chiller, a device that pumps heat when current flows through it, cooled a 0.16-square-millimeter hot spot on a 140-mm2 chip by nearly 15 °C. The researchers say it was the first demonstration of viable chip-scale refrigeration technology.

The prototype brings together two breakthroughs that different thermoelectric groups have hit upon over the past decade. First is the realization that nanoscale layers of thermoelectric material make for much more efficient cooling devices. And second, using thermoelectric heat pumps for cooling the hottest spots on a microprocessor is a much more energy-efficient approach than trying to cool the whole chip.

”For bulk cooling, the rule of thumb is that you have to put in 100 watts [of electric power] to remove 100 watts [of heat],” says Avram Bar-Cohen, a thermal management researcher who chairs the University of Maryland’s mechanical engineering department. But if you apply the thermoelectric material only to a chip’s hottest spots— which represent a small fraction of a chip’s footprint—”you’re only cooling a few watts, so you only have to invest a few,” he says.

The Intel group is the first to demonstrate both concepts in a working chip. They attached a small gallium arsenide substrate to a conventional heat spreader used to cool chips by convection. On the substrate, the researchers grew a 100-micrometer-thick layered structure, called a superlattice, containing bismuth, tellurium, antimony, and selenium. The structure pumps heat from the back side of the chip to the heat spreader. The researchers made a hot spot on the chip with a heat flux of about 1300 W/cm2, much higher than usually found on a microprocessor. They showed that the superlattice caused a roughly 6 °C temperature drop at the hot spot even before the device was powered up, because it conducts heat better than the grease that bonds the rest of the heat spreader to the chip. But once 3 amperes of current were sent through the thermoelectric cooler, the total temperature change was just shy of 15 °C.

”Packaging a nanoscale device within a macroscale system is a big advance,” says Ravi Prasher, one of the Intel researchers who reported the accomplishment in Nature Nanotechnology . But Prasher is hesitant to give a time frame for commercialization because there are still big hurdles to overcome.

Using nanoscale features in a thermoelectric device nearly doubles its efficiency in terms of ZT—the amount of thermal energy wicked away for every unit of current applied. But after a half century of working to improve upon the long-standing 1 ZT benchmark, researchers like Prasher are coming to realize that ZT is not a good system-level parameter. ”It ignores significant issues, such as parasitics—electrical and thermal contact resistance, for example—and by itself is not a guarantee of good performance,” he says.

Lowering thermal contact resistance by an order of magnitude is one of the technical hurdles that must be overcome before a commercial device appears on the market, says Prasher. Limiting the effect of these parasitics is the focus of the team’s research today, he adds.

However, ”starting with a higher intrinsic ZT, allows that much more headroom for parasitics, so substantial cooling is still achieved,” says Rama Venkatasubramanian, a researcher at RTI International who worked on the project.

”This is an important step forward,” says Maryland’s Bar-Cohen, an IEEE Fellow, who was not involved in the research. ”Theoretical analysis of the problem that [University of Maryland researchers] did five years ago said [that] 10 degrees of cooling would make a significant difference,” he says. Not only has the Intel group exceeded that goal, it has also ”demonstrated how this use of thermoelectrics can extend current cooling technology into the next generation,” Bar-Cohen adds. By knocking out the hot spots, he says, you can continue to use conventional air- or water-cooled devices, such as heat sinks, to accommodate hotter-running chips without making the thermal management devices bigger.

Venkatasubramanian says cooling hot spots on processors will be critical to computers in future data centers, ”where computational infrastructure, cooling cost, and energy management are all becoming equally important.”

Bar-Cohen imagines thermoelectrics someday being essential for chips in diminutive portable devices, like multitasking smart phones that must handle simultaneous data-intensive processing tasks. Keeping the devices pocket-size might otherwise require a shift from passive cooling—using the device’s case to radiate heat to the environment—to the use of, say, fans to keep temperatures down and the chip’s reliability up.

Intel’s Prasher agrees but wants to manage expectations. He notes that adding active cooling to a portable will add to the burden on its battery.

To Probe Further

The Nature Nanotechnology article is available here. Thermoelectrics researchers have made several advances over the past year, which were reported at IEEE Spectrum Online, including the development of silicon nanowire-based thermoelectrics, efficiency improvements due to nanostructuring, and better conversion from doping with thallium.

For an overview of thermal management technology, read ”Beat the Heat,” from IEEE Spectrum, May 2004.

The big leap forward in thermoelectric efficiency was reported in Nature in 2001.

The Conversation (0)

The Future of Deep Learning Is Photonic

Computing with light could slash the energy needs of neural networks

10 min read

This computer rendering depicts the pattern on a photonic chip that the author and his colleagues have devised for performing neural-network calculations using light.

Alexander Sludds

Think of the many tasks to which computers are being applied that in the not-so-distant past required human intuition. Computers routinely identify objects in images, transcribe speech, translate between languages, diagnose medical conditions, play complex games, and drive cars.

The technique that has empowered these stunning developments is called deep learning, a term that refers to mathematical models known as artificial neural networks. Deep learning is a subfield of machine learning, a branch of computer science based on fitting complex models to data.

Keep Reading ↓ Show less