Things are getting a bit too hot in the microprocessor world. Again.
Moore’s Law has always come with the caveat that more transistors, switched at a higher frequency, means more heat. Over the years, chipmakers have used tricks like throttling back clock speeds and putting multiple microprocessor cores on a chip to spread out the heat.
But heat continues to stifle chip performance. Hot spots on today’s processors can reach power densities of 1 kilowatt per square centimeter, much higher than the heat inside a rocket nozzle. A growing fraction of transistors on advanced microprocessors are not even operated at any one time because they would generate too much heat, says Avram Bar-Cohen, a program manager at the microsystems technology office of DARPA. “As we put more and more transistors on them, this ‘dark silicon’ fraction has gone from 10 to 20 percent, in some cases more,” he says.
There’s only so much that processor designers can do to keep chips from generating too much heat, and it’s time for some new ways to get that heat out.
The conventional approach to dissipating the heat is to attach the silicon die to a copper or aluminum block carved with elaborate fins and ridges. A plastic fan blows air across the metal. As you can guess, these systems can be bulky, noisy, and power hungry.
Plus, heat sinks and fans won’t cut it for future processors, which will be 3-D stacks of ICs. Such layering can trap heat between chips, making getting rid of it even harder.
Researchers are exploring better ways to cool chips and gadgets, either by redesigning tried-and-tested methods or drastically overhauling them. The challenge for all of these technologies is whether the semiconductor industry will be ready to integrate them, says Ravi Prasher, a director in the energy technologies division at Lawrence Berkeley National Laboratory, in California, who has studied various chip-cooling technologies.
Here’s a look at some methods that could beat the heat.
Tiny Water Pipes
Instead of transferring heat through many millimeters of heat-conducting layers and then into the air, one idea is to put something cold mere micrometers away from the transistors generating heat. This can be done by carving microfluidic channels into a chip or substrate and pumping a fluid coolant through them.
In September 2015, engineers at Georgia Tech demonstrated the first on-chip embedded liquid cooling system on a commercial field-programmable gate array. (FPGAs contain logic blocks that a customer can configure using software to form any computational circuit, with such varied applications as defense, astronomy, and medical imaging.) Electrical and computer engineering professor Muhannad Bakir and his colleagues used conventional microfabrication techniques to etch 200-micrometer-tall channels into the silicon on the back of the chip and connected the channels to external water tubes; the pump and reservoir will eventually be integrated with the chip’s packaging. The system slashed the FPGA’s operating temperature by more than 60 percent compared with that of an air-cooled device. “The technology is compatible with all silicon chips and packaging technologies,” Bakir says.
Another key advantage of liquid cooling: It could be embedded between 3-D–stacked high-power chips. To show this potential, Bakir’s team embedded interconnects into the cooling channels that could be used to link stacked chips.
Fujitsu’s superslim refrigerator is designed to suck heat out of a smartphone. Image: Fujitsu
If water isn’t enough to do the job, some research teams are looking at evaporating refrigerants inside microfluidic channels. By turning into vapor when heated, a coolant can suck out even more heat than a circulating liquid, says Bruno Michel, who leads the advanced thermal packaging group at IBM Zurich Research Laboratory. The vapor turns back to liquid in a condenser outside the chip and returns to repeat the process. Michel says that this two-phase cooling approach, which utilizes low pressure to boost vaporization, should allow them to reach DARPA’s target of cooling chip hot spots by 1 kW/cm2. The tricky part is to regulate the pressure of the system so that the coolant never fully vaporizes. “You have to control the amount of boiling and the vapor quality,” he says.
Simple versions of such two-phase cooling are making their way into smartphones. In March 2015, Fujitsu said it had developed a refrigeration system less than a millimeter thick.
CoolChip has combined the heat sink and fan for added efficiency. Photo: Evan Ackerman
The Fan of Fans
CoolChip Technologies, a startup spun out of MIT, has improved the fan-cooled heat-sink design of old by rolling the two into one. The combination device, which the company calls a kinetic cooling engine, is a flat, circular piece of aluminum carved with fins that spins like a fan. It is half the size of conventional cooling fans, emits much less noise, and can remove 50 percent more heat, according to the company. Microsoft has hired CoolChip to make a system for the Xbox One game console. By early 2016, CoolChip’s technology should be in tens of thousands of gaming-unit cooling systems made by the Taiwanese company Cooler Master.
Thermal greases and polymers are used right now as interface materials for transferring heat from the silicon die to the heat sink. Carbon nanotubes would be excellent substitutes. They have some of the highest-known heat conductivity and are flexible; a rigid material would expand and potentially damage the chip or its package.
The problem is that nanotubes are very stable and don’t easily form chemical bonds that connect them to other parts of the chip package. So “it’s difficult to have good heat transfer between them and heat sources or sinks on either side,” says D. Frank Ogletree, a physicist at Lawrence Berkeley National Laboratory. Ogletree has used organic molecules to form covalent bonds between nanotubes and metal, improving heat flow sixfold. A big remaining hurdle is that only about 5 percent of nanotubes, which are grown in vertical arrays on silicon, touch the metal surface. “If someone could solve this challenge, then it will really be a compelling technology,” says Berkeley Lab’s Prasher.