New Schemes for Powering Processors
Building an on-chip high-voltage transmission grid is one way researchers think they could distribute power better
Corrected 15 October 2009.
14 October 2009—Power grids in the United States are widely thought to need upgrading. The same is starting to seem true of the grids that distribute power on modern microprocessors.
In an effort to keep power consumption reasonable and maintain constant power dissipation, chipmakers have been moving their circuits to lower-voltage requirements, which these days are about 1 volt. But supplying such low voltage to a chip from an off-chip source requires increasing the input current to more than 100 amperes per microprocessor. Carrying such high currents around the chip on copper interconnects leads to high conductive power losses. It also means that the majority of input/output pins—as many as 70 percent—must be devoted to power distribution, leaving few available for transmitting actual data.
Only in the past three to five years has there been much research into how to get power supplies onto the chips for better power distribution, says Eby Friedman, professor of electrical and computer engineering at the University of Rochester, in New York. “If you go to the best companies in the world, they’re just now integrating really coarse power supplies on-chip,” he says.
Tomás Palacios, assistant professor of electrical engineering and computer science at MIT, says that his new technology, ICs made of both silicon and gallium nitride, might help. He bonds together separate wafers of silicon and GaN into one wafer and then processes them, first building silicon circuits in one section of the chip, then etching away the silicon in another section and creating circuits in the GaN below. His idea is to distribute power at a much higher voltage than silicon circuits can stand, perhaps as high as 20 V. Gallium nitride electronics would then convert power to lower voltages and distribute it locally, not unlike the way electrical substations step down high-voltage electricity transmitted from power plants. At high voltage, little current flows, so there is little loss of power in the transmission conductors. And the high-current, low-voltage power travels only locally, which reduces conductive losses and frees up input/output pins.
Friedman says Palacios’s idea seems credible. It’s somewhat analogous to a power distribution idea that Friedman is developing. “He did it at the material level, and I did it at the circuit level,” Friedman says. “You want to make the power supply constant, independent of load. That sounds simple.”
In fact, with a billion transistors on a chip, all pulling current at different magnitudes, different times, and different rates, keeping power constant is challenging. One approach is to use a passive filter consisting of an inductor and a capacitor (an LC filter), which attenuates high-frequency noise. The problem with that setup is that it takes up a lot of valuable real estate on a chip. Friedman’s approach is to make active filters using operational amplifiers. Power supplies based on op-amps take up one-hundredth to one-thousandth the area of those based on LC filters, Friedman says.
Shrinking the power-supply footprint allows chipmakers to put the power supply on the chip itself instead of elsewhere in a system, which increases efficiency and also opens up new possibilities. With the voltage converter off-chip, there’s no ability to supply different loads to different circuits, says Tanay Karnik, a principal engineer at Intel Labs. “You have an option of switching something on or off, but you don’t have the option of running at 0.9 V or 1 V or 1.2 V,” he says. ”If all [the different devices on a chip] have the opportunity to choose their own voltage, you will get a better product.”
Karnik is exploring using a cobalt-based magnetic material as an on-chip inductor to control voltage. The material can be processed like silicon and can handle gigahertz frequencies, as opposed to the megahertz of iron-based magnetic materials. Consequently, cobalt inductors can be made much smaller, once again freeing up chip real estate.
Another approach, being followed by Fred Lee, a professor of electrical and computer engineering at Virginia Tech, is a three-dimensional design. Lee builds the inductors and capacitors on a separate chip, then bonds it to the main chip. This solves space issues but adds the cost of building an extra chip.
In the long run, to handle the demands of a growing number of transistors and to increase power efficiency, experts say that power conversion will have to move onto the silicon chip. Achieving that could require new thinking about the geometry of chip layout to find the most efficient designs. “It will be really helpful if somebody actually has a voltage converter inside silicon,” says Karnik, “but you have to do it right.”
About the Author
Neil Savage writes about technology from Lowell, Mass. In September 2009, he reported on the development of hybrid silicon and gallium nitride ICs.