University of Rochester

Self-Adjusting Chips to Extend Limits of Computing Power

August 8, 2000

A team of scientists at the University of Rochester is undertaking the next step in computing-designing a chip that reconfigures itself as it runs, adapting to the needs of software while processing faster and using less power while doing so. The adaptable chip signals an effort to take full advantage of the massive processing power that chip makers now deliver to desktops every day.

David Albonesi, assistant professor of electrical and computer engineering at the University of Rochester, leads the team, which has created a model called Complexity-Adaptive Processing (CAP) that monitors the way a piece of software uses the microprocessor hardware, and then adapts that hardware accordingly. The result is a more efficient processor that doesn't dawdle while running many tasks. Early tests have shown CAP able to halve the energy consumption of part of the chip while also improving performance.

"Today's microprocessors are pretty inefficient when handling a variety of tasks," says Albonesi. "They're designed to work well overall, but since they're inflexible they can't work as well as they could for any particular program."

The innovation came to Albonesi one Saturday when he decided to lock himself in a room and not come out until he'd thought up something novel. He started to look into certain inefficient parts of a chip, such as the cache, a kind of storage closet on the chip where frequently needed information can be stowed and accessed quickly. Most microprocessors today contain two types of cache, with a larger, slower cache acting as a backup to a smaller, faster one. Although the sizes of these caches are fixed in today's microprocessors, different programs require different sizes to run most efficiently. Similar to how a thermostat controls an air-conditioning system, the CAP design monitors the program as it runs and adjusts the sizes and speeds of the caches as needed, saving the energy taken to maintain them, and saving the time taken to track down information inside them.

Along with Albonesi, Eby Friedman, professor of electrical and computer engineering; Sandhya Dwarkadas, assistant professor of computer science; and Michael Scott, professor of computer science, pooled their resources to develop the system further. The researchers recently received $3 million in funding from the U.S. Defense Advanced Research Projects Agency to continue the work.

The team has a number of other tricks that it expects will produce even greater improvements, including changing the value of "one." Microchips send information by means of "zeros" and "ones," with the zeros represented by no voltage, and the ones represented by a voltage high enough to be detected above the background noise of electricity flowing through the chip. By reducing such things as the cache size, the scientists can lower the overall noise in a particular part of a chip, allowing them to lower the voltage needed to represent a one and thus saving energy. Like the changing cache size, this alteration can be done and undone as needed, millions of times each second, as the processor cranks along.

The CAP model may be able to save even more energy by offering ways to switch fewer transistors in the chip between one and zero, and by slowing down the processor's speed and lowering the voltage when it detects that a program can get by on less. Some of today's processors, such as the Pentium III, have the option of lowering voltage to save battery life on laptops, but this requires running at a slower speed. The team's design should allow even longer battery life while computing just as quickly or faster than today's microchips.

"We're becoming a more and more wireless world, and that means more and more processors draining batteries," says Albonesi. "While computing power has rocketed forward, battery technology hasn't kept pace. By making cell phones and portable computers more efficient, we'll make them run faster while the batteries last longer." Albonesi notes that even non battery-powered computers could benefit, especially for dot-com companies like Amazon that depend on rooms full of power-hungry computers.

Other research teams are experimenting with reconfigurable computing, but using chips that are much slower and don't pack as many transistors. The CAP system, on the other hand, is based on common, commercial chips, and could be integrated into a future version of household PCs without a loss of speed.

"This is leading-edge work," says James E. Smith, professor of electrical and computer engineering at the University of Wisconsin. "This is changing the traditional bounds between hardware and software." Smith says that while brute force powers chips today, streamlining their operations can only make them work better-and someday when current methods have pushed chips as fast as they can go, "We'll have to rely on innovations like this to go faster."