Only a fool would predict the imminent demise of Moore's Law. Thirty-six years ago, Intel co-founder Gordon Moore observed that the density of transistors on chips doubled every 18 to 24 months, doubling the amount of processing power per dollar.
For the past decade, it's been fashionable to assert that Moore's Law is in peril and that microprocessors are straining physics - but you can buy a desktop PC with a 1.3GHz chip for $US900.
At a June conference in Kyoto, Japan, IBM announced that it had altered the structure of silicon in a way that could increase semiconductor speed by 35%.
Strained silicon, as it's called, has been in development for more than a decade and hinges on the tendency of atoms inside compounds to align with one another. For example, IBM placed silicon atop a substrate of silicon and germanium, whose atoms are farther apart than pure silicon's. In response, the atoms in the silicon "strained" about 1% farther apart, according to Philip Wong, a senior manager at IBM Research.
The effect was like enlarging the holes in a sieve: electrons raced through strained silicon 70% faster, for a 35% improvement in real performance.
The real breakthrough is that IBM can manufacture the advanced chips using conventional processes, Wong says. The technology will be used in high-end servers and routers as early as 2003.
Shortly after the Kyoto conference, IBM announced that it had pushed the silicon-germanium technology to speeds of 210GHz in tests while drawing 50% less power than current designs. IBM says that will lead to the production of communications chips running at 100GHz within two years.
IBM wasn't the only company with major news in Kyoto. Intel showed that it has shrunk switches further than many - including Moore - believed possible, unveiling silicon transistors 20 nanometers wide compared with widths of 250 nanometers, which are common today. A human hair is about 60,000 nanometers wide.
The transistors could be used to build chips that run at 1500GHz and draw 1000 times less power than Pentium 4 chips. Energy efficiency is a bonus property of smaller, faster transistors and will extend the battery life of mobile devices and reduce heat buildup. Intel says the transistors can be built using today's processes and will be in Intel chips in 2007.
Impressive advances notwithstanding, silicon will one day run out of steam. You can only cram so many switches on a chip before electrical leakage, heat buildup and manufacturability present insurmountable obstacles. The need for increasingly precise tooling will drive the cost of a semiconductor fabrication facility to $US200 billion by 2015. That kind of price tag prompts companies to build better mousetraps. IBM, Intel and others have postponed the day CMOS technology becomes obsolete, but that day will come.
One promising successor is the carbon nanotube, which draws its name from its material (pure carbon), its size (1 to 3 nanometers in diameter) and its shape (a tube composed of hexagonal structures). Nanotubes are light yet strong, conduct heat well and are better electrical conductors than copper.
Nanotubes have astonishing potential, but making them work together is difficult. "We can make an array of [nanoscale] structures," says Paras Prasad, director of the University at Buffalo's Institute for Lasers, Photonics and Biophotonics in New York. "The challenge is making the electronic and photonic connections."
Some scientists foresee production nanotube chips 16 times faster than today's as early as 2004. But others disagree. "Organising [nanotubes] to form a circuit is nowhere near application," says Dmitri Antoniadas, director of the Microelectronics Advanced Research Center at MIT. "For that, we're probably looking beyond 2015."
In 1999, researchers at Yale University and Rice University created both switches and memory elements by altering organic molecules so they would trap electrons at given voltages. They've shown promise as reusable switches that can, given the right conditions, self-assemble into a desired configuration, essentially manufacturing themselves at a fraction of the cost of building conventional transistors. Such chips could be 1000 times faster than current technology, and memory speed and density could be increased by an order of magnitude.
Ulfelder is a freelance writer in Southboro, Massachusetts. Send email to Steve Ulfelder.