William Pohlman's 1999 retirement party, capping a 20-year tenure at Intel, had barely started when two former colleagues pulled him aside. They wanted him to help them start a new chip company, Primarion, which wound up focusing on regulating the energy demands of microprocessors. The back nine could wait. "I knew the technology megatrends that were creating problems for building future chips," says Pohlman, a former vice president of engineering in Intel's microprocessor group.
Two years later Intel (nasdaq: INTC - news - people) is pulling off breathtaking feats, squeezing 42 million transistors onto a sliver of silicon 217 square millimeters. Its new Pentium 4 churns through data at rates of 1.5 gigahertz; one gigahertz is a billion clock cycles a second.
But as chips get this dense and quick, they get hot�hot enough to boil water. The heat makes them so electrically "noisy" that they can fail. And the materials that enabled such chip density are reaching physical limits. Says Pohlman, now chairman of Primarion in Tempe, Ariz.: "We've hit an inflection point."
And it's a biggie. If nothing is done to rethink chip design, the most powerful microprocessors could be consuming more than 1,000 watts by 2004. "If it's business as usual, we wind up frying eggs" with microprocessors, says Dennis Monticelli, a Fellow with National Semiconductor (nyse: NSM - news - people) in Santa Clara, Calif.
Many of these problems could occur within two chip generations, about four years from now. Since it takes about two years and more than $1.5billion to build a new semiconductor factory, chipmakers are rolling some expensive dice betting researchers will find solutions in time. "Our goal is to make Moore's Law work for the next decade," says Patrick Gelsinger, chief technology officer at Intel, referring to the tenet that the number of transistors on a chip doubles every 24 months.
Doing so, however, will demand changes in the design of chips and the materials that compose them. At a recent industry conference, Gelsinger declared that managing heat is now one of the industry's top challenges. Switching every transistor off or on requires a touch of energy. As transistors shrink, it becomes impossible to completely turn them off, so they leak current all the time. That draws electricity�50 watts in the case of the Pentium 4�and the electricity creates heat. Without the use of cooling techniques, temperature spikes above 105 degrees centigrade have occurred. Piping out the heat is expensive. Even simple "heat sinks," chunks of material that carry heat away from microprocessors, can add $16 to the cost of a $600 chip. More elaborate models, with tiny chambers of water that vaporize and carry away heat even more quickly, can run twice the cost. Computer makers hate adding gizmos to their boxes to flush out heat, preferring to save the room for gear that makes their machines more appealing.
New cooling tricks are starting to emerge. In late February the fledgling Incep Technologies in San Diego introduced a technique for packaging together a microprocessor, a logic board for regulating power to the chip and a heat sink. Even though such "encapsulation" could cost $200 per unit, Incep President James Kaskade contends that it both cools the chips and saves space inside the box.
Isonics Corp. (nasdaq: ISON - news - people), in Golden, Colo., a maker of specialty materials and chemicals, is proposing a new material:a "purer" version of silicon called Si-28, which channels out heat better than conventional silicon. The silicon in typical wafers is a blend of three silicon isotopes. Sifted down to just the Si28 isotope, Isonics' wafer conducts heat better.
Even though Si28's thermal properties are attractive, changing materials could be an expensive option, adding at least 25% to the cost of the wafer. Isonics Chief Executive James Alexander says he needs committed partners before manufacturing the first wafers. He claims that Advanced Micro Devices (nyse: AMD - news - people), among others, is experimenting with the materials.
Even better than getting the heat out would be generating less of it in the first place. Intel's Gelsinger is exploiting several tricks to make chips more efficient. Adding more local memory, or "cache," to a chip reduces the work the microprocessor must do to fetch needed data. Letting two microprocessing units share one cache cuts work even further. Designating a special section of the chip to handle common tasks also helps. So, too, does handling repetitive tasks together.
Both Intel and AMD are also trying to be smarter about how their chips use power by using software to deliver just enough juice to the chip to get a job done. "The chance that you need the highest performance at any one time is small," points out Frederick Weber, vice president of design engineering at AMD in Sunnyvale, Calif. Instead, chips might operate at clock speeds ranging from 300 megahertz to 1,500 megahertz, depending on the tasks.
Transmeta (nasdaq: TMTA - news - people), a much-talked-about Santa Clara newcomer, is taking a different approach entirely. Instead of slowing down a fast processor, it is using software to replace transistors. Transmeta's technique, called "code morphing," translates the instructions sent to a chip into bigger chunks that can be handled more efficiently. The result: Its Crusoe chip, which uses about 1 million logic devices such as transistors, is already used in Sony and Hitachi laptops. Transmeta and its competitors argue about whose chip performs at what speed. "Racing for megahertz isn't the goal�giving consumers a great experience is," says the company's founder and chief technology officer David Ditzel.
A more insidious problem, the one that lured Bill Pohlman out of retirement, is the dreaded power spike. Operating at gigahertz speeds takes a lot of energy, so designers must lower the voltage they apply to transistors so as not to fry the electronics. But at lower voltages the signal that pulses through the chip gets so weak it could get lost in the chip's electrical cacophony. Imagine 50 million doors slamming every fraction of a second. And, when the electric potential dips below one volt, devices may not get enough juice to switch. A power-hungry transistor will steal energy from its neighbors, causing a tiny surge on the chip. "Either you have to run your processor slower, or you could get a �blue screen'�the system fails," says Pohlman.
He thinks he has an answer to these concerns by judiciously managing the voltage. Primarion is designing small, special-purpose silicon germanium chips that sit next to a microprocessor, monitor its energy demands and supply the right amount of power at the right time. "We think it might add $20 to the cost of the microprocessor but it could run as much as 20% faster," Pohlman argues. Primarion's first chips, which operate about five times as fast as the top microprocessors, might be ready by year-end. (Silicon germanium chips run so fast because electrons travel more easily through the material than they do through silicon.)
As transistors get even smaller the materials that have been so reliable for chip designers begin to give out. One standard ingredient has been silicon dioxide, a combination of silicon and oxygen atoms that makes up beach sand and quartz crystals. Silicon dioxide has played two different roles for transistors: It insulates the tiny metal wires connecting those millions of transistors and manages the process of turning a transistor's power off and on, serving as a buffer between positive and negative charges. By thinning this "dielectric layer," designers have sped up transistor-switching. But it will soon be stretched about as thin as it can go: The silicon dioxide layer on the daughter of the Pentium 4 will be a mere six atoms thick. Designers can't scrape away too many more atoms or else those lines will touch or interfere, garbling the digital signals.
Researchers despair of ever finding another material that can both manage the switch and insulate the wires. That leads them in different directions: adding new materials to the dielectric material governing the switch and trying to concoct new insulators for the wires. IBM and others are trying a grocery list of materials. In early March, for instance, Dow Chemical (nyse: DOW - news - people) unveiled a porous organic material that it promised to make available as an insulator later this year. One radical idea for insulating the wires would be to leave nothing but air between them, says Daniel Dawson, a manager at IBM's Almaden Research Center. Such a chip might be too fragile, however.
Many solutions are under way, but if the biggest chipmakers don't settle on an approach, it will be difficult to drive down the costs of future chips. One compromise: throwing in a pair of oven mitts with every new computer.
Getting The Heat Out
| To make more powerful microprocessors, engineers try to squeeze more transistors onto a single silicon chip. That means transistors have become vanishingly small. If Intel's top–of–the–line Pentium 4 processor measured 500 miles on a side, then each of its 42 million transistors would be only 19 feet across the top. But the tinier the transistors, the hotter the whole chip becomes. Here are a few of the techniques designers are trying to get the heat out. Just about all of these, however, add some cost and difficulty to chipmaking. | ||
| FIND A BETTER WAY TO FLUSH OUT HEAT | USE DIFFERENT MATERIALS | REDESIGN THE CHIP |
| 1. Heat sinks. These are chunks of material that pull heat away from the microprocessor. Metal is a good conductor of heat, water is better. (Air is the best.) Some designers are building novel heat sinks with tiny water chambers. The water draws out the heat, vaporizes and, as it cools, condenses again. 2. "Encapsulate" a microprocessor. Startup Incep hopes to package a chip, heat sink and the ability to modulate voltage. By sliding the chip into this tidy package, it could expose more of it to air. | Some firms are exploring how different materials flush out heat or switch faster with less voltage. 1. IBM's "silicon on insulator" layers silicon and an insulator such as silicon dioxide where current passes through the semiconductor. 2. Materials–maker Isonics pushes the idea of using isotopically pure silicon wafers that have fewer crystal defects than conventional wafers. Electrons pass through with fewer road bumps, generating less heat. Isonics is seeking a commitment from a big chipmaker before it begins manufacturing such wafers. Chipmakers are nervous about the cost. | 1. Add more local or "cache" memory. To do a task such as addition, a chip might have to fetch the numbers from the hard drive. Adding more cache memory means the chip does less work to find the data. 2. Add specialized processing blocks. Creating a section of the chip designed to handle repetitive tasks efficiently saves work. 3. Add another processor, but share the memory. Creating two processing units that share a large cache on the same swatch of silicon speeds up work. 4. Use software to find parallel tasks. In the most radical case, software might be able to reorganize a problem so that a chip can handle several tasks simultaneously. Software can also let a processor work as if it were two units when it's only one. -E.C. |
 | Feeding the Pentium Beast 04.02.01 from Too Hot to Handle Intel's astonishing march toward ever denser chips comes with a cost: skyrocketing energy demands. The prospect of 100-kilowatt chips has designers scrambling for solutions.
Projection figures assume no advances in energy efficiency techniques. 1Leakage is the dissipation of energy as a result of imperfect transistor function. Source: Intel. |
http://www.forbes.com//forbes/2001/0402/126.html
