Nanometres are smaller than you can possibly imagine. They're as much smaller than a metre as a marble is smaller than the Earth. Yet the success of the computer hardware industry depends on regularly shaving off a few tens of nanometres every couple of years: this year, we're due to see chips based on 90nm rules begin to take over from the current 130nm processes that fuel the game.
Switching processes is an expensive game, and with each shrink the price of entry goes up. Once everything's running, you make more money from a smaller process than a larger -- you can fit more circuits onto a single wafer, reducing the per-unit cost; your circuits run faster and can have more components, which gives you more capabilities. The ideal is to find some new market that just wasn't possible under an old technology -- but frequently, you don't know whether it's there until you can produce the chips. Next to exploring for oil, it's one of the biggest gambles you can make in business.
New processes have two big hurdles to pass before they start to repay a company's investment. First, they have to be shown to work: here, 90nm is fine. People have paraded their pre-production samples at trade shows, to their customers and to analysts. Then, they have to be made in commercial amounts.
So much can go wrong. On smaller chips, the distance between components -- and between features in those components -- is necessarily shrunk. Wires that run alongside each other interfere with each other's signals, and this crosstalk gets worse the closer they are together. Smaller conductors have greater electrical resistance, smaller insulators have less: leakage currents go up while working currents go down. Noise becomes a bigger problem with smaller voltages. And all this is entirely independent of the sheer physical problems of mass-producing circuits where the active components are getting close to the atomic realm: brand-new technologies are needed for every step. Even the design tools for creating, laying out and simulating the circuits need to be fundamentally changed as different aspects of semiconductor physics start to become important.
It's a huge experiment, and one where you only know how well your production works after you've shelled out the billions of dollars to build the plant. Rumours say that actually making 90nm chips is proving harder than expected, and that's a problem. Everyone is acting as if 90nm is the only place to go. Intel's next big steps -- the Prescott Pentium successor, the Montecito Itanium upgrade, the high performance Centrino Pentium Ms -- will all be 90nm, and they're all late. AMD's launched its Opteron at 130nm, but the future is very much 90: it's here where there's room to make performance gains that really differentiate a product. Has AMD actually got 90nm chips? No. Coming soon.
You can tell when production isn't following schedule, and not just because the chips aren't around. The original Pentium was fashioned out of a positively obese 800-nanometre process. At the time, though, that was cutting edge and it almost didn't work. Intel finally launched the processor at two speeds a hairsbreadth apart, 60MHz and 66MHz. That's always a sign that the production line making enough marginal parts that it's worth shipping them at a slower speed than you'd like: watch out for more odd speeds this year.
Of course, Intel successfully wrestled the problems with the 800nm process to the ground -- but that's not been true of every good idea. Gallium arsenide was once thought likely to dominate high-speed computing, being the semiconductor of choice for things such as Cray's series of world-beating machines, and lots of people spent lots of time building production lines for it. In the end, economics meant that boring old silicon kept being just good enough for even the high performance stuff: you'll still find GaAs circuits in high speed communications devices -- a big niche, but a niche. There's no guarantee that your gamble will pay off.
There are some 90nm chips on the market. Xylinx has a programmable logic chip, the Spartan 3, out -- although some of the key electrical characteristics are still being assessed by the company, another sign that production isn't being quite as predictable as it should. Even the previous generation of fabs, 130nm, had and has problems -- the transition from 250nm to 180nm went very smoothly, and companies weren't expecting the range and tenacity of problems they got when going to 130nm. For a long time, many lines were running at yields of 20 to 50 percent -- figures where the economics only just work. You need 70 percent and better to really take advantage of mass production breaks.
You'll not get any discussion of yield out of chip companies, who guard this information with the same jealous paranoia that Bin Laden reserves for his mobile phone number. But whenever a chip company executive bears down on you waving roadmaps, presentations and glossy predictions, remember this: until they've actually made the chips, they don't know whether they can. That's never been more true than with 90nm, where small issues may be big problems by the end of the year.