In 1989, the state-of-the-art Intel 486DX processor dissipated 4 watts, prompting PC makers to moan about adding a heat sink. Fourteen years later, Intel's newest Pentium 4, code-named Prescott, burns well over 100 W.

During the past decade, power dissipation has become the most important factor in processor performance, influencing every decision-from architecture to circuit design. Desktop, laptop and embedded processors are all constrained by power budgets that are at or near physical limits.

Supply voltage has been the main weapon to fight rising power. But CMOS transistors stop operating below 0.8 volt. With supply voltages now at 1.2 V and falling, that approach is nearing its limit as well.

Within a few years, we may reach the point where supply voltage and power budgets can no longer be improved. At that point, performance gains from process shrinks will be constrained to perhaps 10 percent per year, a far cry from the 50 percent or so we have come to expect. Why build smaller transistors if we can't afford to run them faster?

Alternative approaches are needed. Embedded processors, for example, have used one or more simpler processors to achieve high performance with modest power.

But greater savings can be achieved by shifting common operations to fixed-function logic. For example, Alacritech's TCP offload engine can execute the TCP stack at 1 Gbit/second using just 1.5 W, a task that would require at least 30 W on a Pentium processor.

The biggest power waste in high-performance processors is leakage current. As transistors become physically smaller, electrons more easily leak from one side to the other, even when the gate is off.

Also, transistors designed to switch quickly leak more than do slower transistors. Leakage current can waste half of a chip's power in 90-nm technology, and the problem will get worse in the future.

Recently, Transmeta Corp. announced a new technique that dynamically adjusts leakage current, allowing more leakage when the processor needs to run faster and cutting back when it is not busy. This type of fundamental advance will be needed to increase future processor performance. Simply slapping the design into a new manufacturing process won't be much help.

http://www.eet.com