Chipmakers have been demonstrating technologies they believe will deliver processors clocking more than a thousand times faster than today's leading-edge products within a decade.

The timeframe is in line with Intel co-founder Gordon Moore's 1965 prediction that processor computing power - roughly the number of transistors per chip - would double every 18 months. But many people expected this so-called 'law' to break down sooner.

The problems are formidable as clock speeds push past one terahertz (1,000GHz) and developers reach the limits of miniaturisation, addressing scales at the level of a single atom. You need to look no further than the Pentium 4 with its massive cooling needs to see that one of the more obvious issues is heat.

What's cooking?

The electrical power density of a P4 - that is, the watts dissipated per square centimetre - is roughly that of the filament of an electric cooker. Intel architecture chief Pat Gelsinger reckons that, if today's transistors were to be scaled down year-by-year to satisfy Moore's Law, their power density would reach that of a nuclear reactor by 2005, and of the surface of the sun by the end of the decade.

Clearly, then, transistors have to change. Gerald Marcyk, Intel's director of component research, said: "Simply making transistors small is not good enough any more. We need to make them small, fast and efficient." Intel's immediate goal is to make processors with 25 times the number of transistors in today's P4, running at 10 times the speed but using the same power.

Transistor structure

Early chips had a few hundred transistors, the P4 has 32 million, and by 2007 processors are expected to have as many as a billion. You might think this huge increase alone could account for the heat rise, with each transistor requiring its little quota of power. But it is not as simple as that and, to see why, you have to delve into the microstructure of transistors.

A digital transistor (as opposed to one used as an amplifier) is just an on-off switch. The basic Complementary Metal Oxide Semiconductor (CMOS) transistor consists of three elements: a source, a drain and a gate that is insulated from the other two by a thin layer of silicon dioxide. A drive voltage is applied across the source and drain and, ideally, current flows between them only when a suitable voltage is applied to the gate.

The good news, as the structure gets smaller, is that the drive voltage can drop, resulting in a proportionally larger drop in heat generated because power is a function of the voltage, squared. This is why the number of transistors on a chip does not necessarily indicate its power drain. Early, so-called Transistor-Transistor Logic worked at 5v; the latest mainstream processors are pushing towards 1v.

This trend does not continue indefinitely. Other factors become significant as transistor dimensions become vanishingly small. The resistance of the tiny electronic paths soars, requiring a higher voltage to maintain the current, in the same way that it is easier to produce a given flow of water through a large pipe than a small one.

Current leakage

There is also the matter of leakage currents, which flow across the gate's imperfect oxide insulator, and between the source and drain even when the transistor is off. It is these that cause the power drain when a processor is at rest.

As transistors get very small, with their insulating oxide perhaps only three atoms thick, leakage can rise very steeply. The result is an exponential rise in power consumption.

Tiny transistors can even be affected by the cosmic rays that constantly bombard the Earth. An atom that gets ionised by being struck by an alpha particle may not matter if it is one among a large number, but it can change the logical state of a very tiny transistor. Such so-called soft errors are a common problem in memory cells.

Water, water everywhere

More difficult to understand is the effect of capacitance, which is a measure of how much charge something can hold. Again, the best analogy is with water. Think of a river feeding a reservoir before continuing to the sea: only when the reservoir is full will any water start on the final leg of the journey. In the same way, a transistor will only pass the current needed to trigger itself on once it is charged up.

Not a problem, you might think, as transistors get tinier. But remember that a terahertz transistor needs the agility to empty and fill itself a thousand billion times a second. The tiniest capacitance, causing what is called gate delay, can be significant.

The smaller the gate delay and the higher the drive current, the faster the transistor. To this end designers seek to reduce the resistance and capacitance.

One answer, pioneered by IBM, is the Silicon on Insulator (SOI) transistor. The source and drain sit on a very thin connecting layer of silicon, which itself lies on an oxide insulating it from the silicon substrate. This considerably reduces both the source/drain leakage and the capacitance.

Emerging designs

Late last year, at around the time of the annual International Electronic Devices Meeting, there was a flurry of announcements on emerging chip designs from different manufacturers.

IBM said it had created a working double-gate transistor, not a new idea but one which many believed would not be practical. The gates sit above and below the semi-conducting channel separating the source and drain, and allowing it to be thicker and carry more current.

Intel, which had long played down the importance of SOI, did a volte-face and showed what it said was an improved version. In this, the source and drain sit directly on the insulator, leaving a tiny semi-conducting channel between them and, according to Intel, considerably reducing leakage. Intel has also thickened the source and drain round the lip of the gate to reduce the resistance.

Finally, Intel has replaced the gate's silicon dioxide insulation with zirconium dioxide, which allows the layer to be thicker without a rise in capacitance and is said to reduce gate leakage by a factor of 10,000.

Body of evidence

Intel claims that the design is also likely to reduce soft errors and will eliminate what is known as the 'floating body effect', a tendency for charges to float in the substrate and alter the device's behaviour. Elements of the design will be introduced into Intel chips as soon as 2005, and by 2010 it will allow chips to operate at only 0.6v.

Intel says it has already produced a transistor with a 15nm gate running at 2.6THz at 0.8v. The company has yet to show that it can integrate such devices into a processor and mass produce it. But Marcyk said: "We have shown that we can produce the technology for devices that will go into production by 2009. That's very good news."

Work in progress

The shine seemed to have been taken off Intel's announcement when AMD demonstrated a conventional CMOS transistor with a 15nm gate clocking more than 3THz. However, it was not clear how useful the device would be in practice, although the fact that AMD is not abandoning work on SOI designs may signal that it is less than ideal.

An AMD spokesman said: "It shows that there is life in the old CMOS transistor yet."