Logo IMG
HOME > PAST ISSUE > May-June 2006 > Article Detail


Reverse Engineering

backward and forward both run to need may they ,faster run to are computers If

Brian Hayes

Computational Fuel Economy

Why is power consumption so important, and what does it have to do with reversibility? The chain of reasoning that links these concepts is a fairly long and tangled one, but the individual steps are easy enough to follow.

In the past two decades the performance of microprocessors has improved by a factor of 1,000, but the trends that have made computers more powerful have also made them more power-hungry. Some chips dissipate more than 100 watts and require elaborate fans, heat sinks and even liquid cooling. Designers would welcome another thousandfold gain in performance, but they cannot cope with any further increase in power density. Single chips that consume electricity by the kilowatt are just not a practical option.

Where does all the energy go? Much power is lost because neither conductors nor insulators are perfect; electrons meet resistance where they ought to pass unopposed, and they leak through materials where current ought to be blocked. Both of these problems will get more severe as silicon devices continue to shrink. Another energy drain is the need to accelerate and decelerate electric charges as signals move through the circuitry; this cost goes up along with processor speed. And it always takes energy to push electrons "uphill" against a voltage gradient.

Some of the strategies for reducing the energy demands of a computer are much like measures to improve the fuel economy of an automobile. To get better gas mileage, you make a car lighter and aerodynamically sleeker; likewise in digital circuits, you can reduce inertia by using fewer electrons to represent each bit of information, and you can cut resistive losses with better conductors. In the car, you drive slower and more smoothly; in the computer you operate at lower voltage and avoid abrupt swings in voltage. For even greater savings in an automobile, you might try a hybrid design, with a battery or a flywheel to recapture energy invested in acceleration and hill-climbing; the electronic counterpart is an experimental technology called charge recovery.

In the world of chipmaking, some of these energy-conserving measures are already well-established tools, and others are likely to be adopted soon. For example, copper is replacing aluminum in the metal interconnections on some chips to improve conductivity. The voltage levels of on-chip signals have fallen from 5 volts to as little as 1 volt. Further steps of the same general kind may well avert a silicon energy crisis for another decade or two. But then what? If the number of operations per second is to increase by a factor of 1,000 without raising power consumption, then the average energy per operation must be reduced to a thousandth of its present value. Is that possible, even in principle? What about a millionth?

comments powered by Disqus


Subscribe to American Scientist