Mixed signals in semiconductors


Statistics abound in the world of chipmaking. Some of the most useful are those prepared by Semiconductor International Capacity Statistics (SICAS), a body set up by the biggest chipmakers to track how much of the world's production capacity is being used. With its copious graphs and tables, the SICAS quarterly report provides a window into the volatile world of silicon production.

The latest figures, for the first quarter of 2006, show that the current boom in the sector differs from previous upswings in some interesting ways. They tend to indicate that the market is not overheating and that the current growth spurt has some way to go. However, they also provide support for those who believe that the latest round of upgrades to the sales estimates for 2006 were misplaced.

Encouraged by strong sales in the year so far, the World Semiconductor Trade Statistics (WSTS) organisation upped its forecast for the year. Instead of sales growing 8 per cent during 2006, WSTS now thinks they will increase by just over 10 per cent. In contrast, analysts at JP Morgan thought that a slip in the April figures, largely due to slow sales of PC microprocessors, indicated trouble ahead. The firm cut its forecast for 2006 from 11 per cent to 9 per cent, believing that the market has hit its peak.

The latest SICAS figures do not show what happened in April, but they do indicate a slowdown was underway during the first quarter. They also show that the usage of fabs in the last few quarters has been much more uneven than is usually the case in a strong market.

SICAS splits its figures into groups of processes. The one that most people watch is the leading-edge bucket. For SICAS, that means anything with minimum feature sizes below 0.12µm. That means the 0.11µ, 90nm and the brand new 65nm processes. Typically, it's unusual to see these at anything less than 90 per cent of capacity. Fab owners do not like to invest lots of money in equipment to see it go idle, so they try to track the industry as much as possible.

Fab owners will normally increase capacity in response to upswings and, because of the delays in getting extra capacity online, they normally lag actual demand. At the same time, the companies selling chips like to take advantage of the lower cost offered by doubling the density of circuitry on a piece of silicon that the move to each successive generation offers. So, they queue up to get on the new processes. The result? Capacity is pretty much sold out unless the market is in poor shape. It will be no surprise that utilisation of leading-edge processes was at its worst from the start of 2001 until late 2002.

Even in the downturns of 1996 and 1998, leading-edge utilisation only narrowly dipped below 90 per cent. In the bleak years after the collapse of the Internet bubble, utilisation sank close to 80 per cent. But that was still a lot better than what happened to older processes, all of which saw utilisation slump to less than 70 per cent and, in most cases, less than 60 per cent.

In an upswing, utilisation tends to converge in the high 90s. You can normally tell when this market is overheating when just about everything is more or less sold out. You could see that happening in 1995, 1999-2000 and, most recently, in the first half of 2004. The last cycle ended when companies noticed they had built up way too much inventory and cut back on buying chips. So utilisation of everything, bar the leading-edge stuff, quickly dropped back to the mid 80s, slowly climbing back during 2005.

Although one or two foundries have remarked on how they have sold out of capacity on some of the older processes, the industry in general has not seen the same effect. And, in the latest quarter, after utilisation nudged 90 per cent for most proceses, the number slipped back to the mid 80s again. Thanks to the lull in microprocessor demand, leading-edge utilisation also took a slight knock, dropping two percentage points.

If that situation continues it means that chipmakers will find it difficult to keep prices up, which will dampen growth. The move to 65nm may give chipmakers at the high end the opportunity to charge more per chip, which will push the average higher, but the effects of that move will not be seen for some months. And the big buyers are good at watching the markets to see who is vulnerable to a bit of price pressure. Unless demand begins to approach supply, the bears will have a better chance of being right than those who just upped their forecast for 2006.