Looking back at the history of semiconductor process technology, it's easy to see why we've become complacent about the pace of our progress. Every two years or so, we get a new technology generation that doubles the number of transistors on a chip of constant size; while making smaller advances along the way, we stay roughly on the Moore's Law curve.

Bernie Meyerson, vice president and chief technologist at IBM, calls this "turning the 'small' knob." Every so often, process engineers reach over and give the knob another turn, making chips smaller, cheaper and faster. It hasn't always been this easy, and it's about to become a lot more difficult.

Until the early 1990s, computers were built out of integrated circuits made with bipolar transistors. These bipolar ICs consumed power whether they were operating or not, but in the early days, the relatively low numbers of transistors per chip and per system made this power consumption tolerable. As complexity increased along the Moore's Law curve, bipolar technology's power consumption began to get out of hand.

CMOS logic was developed to solve this problem. The CMOS chips of the 1990s consumed almost no power, except when transistors were active—switching a signal on or off. Apart from the trivial standby current when the chip was inactive, power consumption was low and proportional to performance.

With each new generation of CMOS process technology, leakage became less of a problem. Today, leakage is getting out of hand. Active power consumption is growing 30 percent per generation while standby power triples. Not only does leakage threaten to become the largest factor in a chip's power consumption but it can also vary widely from wafer to wafer and from chip to chip. Previously insignificant variances in manufacturing tolerances can turn a chip designed for use in a cell phone into one better suited for a space heater.

What technology threatens, technology can also nurture.

What technology threatens, technology can also nurture. An alphabet soup of new manufacturing processes and design techniques is being developed to reduce leakage current. At Microprocessor Forum 2003, Transmeta announced that it's developing a way to vary the threshold voltage (Vt)—the point at which transistors switch on or off—of its processors, possibly hundreds of times per second, according to the needs of the programs running on the system at each moment. If a user needs speed, the system can run faster and hotter. But when the system sees that the user is just typing an e-mail message, it can reduce clock speed, voltage and leakage current to the lowest-possible levels.

Intel got a lot of publicity in November, when it announced that it had achieved a "breakthrough" by redesigning the basic CMOS transistor to use a new insulating material and a metal gate that replaces the standard silicon gate material. Although perhaps a breakthrough for Intel, this configuration has been widely developed and described over the last several years by IBM and other companies. In the view of IBM's Meyerson, more research is needed before a specific transistor configuration can be chosen for the 45nm process generation, likely to go into production in 2007 or so.

Each generation requires a greater number of fundamental new advances in semiconductor technology. Just to enforce Moore's Law for the next 10 years, we'll have to change the very nature of transistors, not just their component materials. Today's production is the work of engineers. Tomorrow's processes are the work of scientists. Beyond that is the realm of science fiction.

How easily will we knock the leakage problem? Send your thoughts tofeedback@eb.reedbusiness.com.

Peter N. Glaskowsky (chipadvisor@ideaphile.com) is editor in chief of In-Stat/MDR's Microprocessor Report. For more information on topics covered in this column, visit www.chipadvisor.com.