Anthony Doesburg: Silicon revolution is running out of breath

If you've upgraded your PC lately and were disappointed that it failed to blow its predecessor out of the water, there's no point in taking it back to the shop. The sad fact is that the PC performance growth spurts we've come to expect every couple of years are harder and harder for manufacturers to deliver.

The problem is physics. Microprocessor designers at Intel and elsewhere are running into the limitations of silicon, the element from which computer chips are made and on which the information technology revolution rests.

As they pack chips with more and more transistors, silicon's hitherto predictable electrical conductivity is going haywire.

For the past few decades, chip-makers have merrily followed Moore's law, which predicted the doubling of transistor density every couple of years. But what was only ever a rule of thumb put forward by Intel co-founder Gordon Moore is beginning to come up against the laws of nature.

"The ability to keep fulfilling Moore's Law ... is clearly causing the semiconductor industry a few sleepless nights," says Martin Allen, of Canterbury University's electrical and computer engineering department.

Instead of being able to cram ever more transistors on to a silicon chip - Intel's 2008 vintage i7, to be found in many new PCs, has 731 million - electronics engineers are having to find other ways to meet the need for speed. One approach is through microprocessor design and another is to look at alternatives to silicon.

First, though, why the obsession with transistors? The theory is that the more you have, the greater a chip's processing capacity.

A transistor can be thought of as a gate that can be either open or closed, those states corresponding with the zeroes and ones of a computer program. The greater the number of transistors, the faster the chip can chomp through the program's instructions.

The theory breaks down when transistors are shrunk too far, says Allen. The transistors on Intel's first i7 chips were 45 nanometres long, about a 1000th of the width of a human hair. New i7's are 32nm, and the industry is heading for 16nm transistors.

That's the realm of nanotechnology, at which the usual rules of physics give way to quantum mechanics. Electrons, which reliably pass through the gate of a 65nm transistor, start behaving unpredictably at smaller sizes, Allen says, resulting in data loss.

"As you get into these smaller and smaller dimensions things become weirder and weirder. You're dealing more with probabilities than absolutes, which is no good when you're storing data - you want to know whether you have a one or a zero in a particular location, not that it could be a one or could be a zero."

The electron tunnelling effect that rears its head below 65nm is a property of silicon. Researchers are experimenting with adding other elements - hafnium, for instance - to silicon to overcome the limitation, or trying to develop silicon replacements using oxide-based semiconductors.

"Oxide-based semiconductors have very attractive properties but we're still trying to understand how they work," says Allen.

"There's nothing on the drawing board yet."

According to a paper in the March 26 edition of the journal Science, the research is at a similar stage to where semiconductor technology was 50 years ago and, while the challenges of managing the interfaces between complex oxides are enormous, so is the technology's potential.

"Today we are in the midst of learning how to meet this challenge; once mastered, these interfaces will provide vast and unforeseen opportunities, technologies, and science for decades to come," says the paper, entitled "Oxide Interfaces - An Opportunity for Electronics".

But silicon is going to be a hard act to follow, Allen says. "Silicon is still going to be around for a very long time because it's the perfect material. It's so easy to process, it's so easy to put these millions of transistors on."

Having gone almost as far as silicon miniaturisation allows, chip-makers are taking a brute force approach to increasing processor performance.

The way in which Intel manages to get 731 million transistors into one of its i7 chips is effectively by assembling it out of four processors, or cores, and earlier this month it was reported the company had developed a 48-core chip.

But adding cores is like throwing more manpower at a job: it increases output but at the cost of greater energy consumption and waste heat generation.

Getting the cores to work together better is the role of software, which Allen thinks could be the answer to keeping Moore's law intact until the new era of electronics arrives.

"Rather than just rely on the pure grunt of these things you start to become a bit more clever about how you use them."

Software writers could be about to have their day.