03.27.05
High-tech boom and bust cycles
Twenty years ago, my then-boss Jim Mikkelson gave me a compelling explanation of why the semiconductor industry has such impressive boom and bust cycles. It has to do with the semiconductor manufacturing process being incredibly sensitive and the design cycles being short and somewhat predictable.Imagine that the semiconductor industry has just started a boom cycle. Orders at the fabs — semiconductor manufacturing plants — has increased dramatically and suddenly (for reasons that I’ll explain later). This means that the fab increases its production. (In fab lingo, the wafer starts — the number of wafers put into the beginning step — goes up.)
Positive feedback loops: quantity vs. yield
Unfortunately, the fab process is incredibly sensitive to changes in the process. A small change can cause the yield — the percent of chips on a wafer that turn out to function correctly — to go down significantly. (At the fab I worked at, we once lost an entire month’s worth of product because a wafer cleaner was miscalibrated.)
One common response is to start even more wafers through the fab. You might think this is counterproductive, but usually the yield doesn’t go down so much that starting more wafers can’t fix it, it takes a while to diagnose yield problems, and the customers want their chips NOW. Unfortunately, running more material through the fab usually makes yield go down.
Hiring more people and buying more equipment also usually reduces the yield, at least in the short term.
Just as the fab starts to get its yield problems under control and get the new equipment up and running, orders suddenly plummet dramatically (for reasons I’ll explain in a minute). So the fab cuts back on its wafer starts, and that starts a positive feedback loop where the yield keeps going up.
Chip customers
At this point in the cycle, chips are incredibly plentiful and, because the fabs have excess capacity and high yield, ridiculously cheap. At this point in the cycle, Jane Engineer notices that and thinks, “Hmmm, with one of those and four of those, I can make a doohickey that is twice as good as what the competition has and at a quarter of the cost. We could take twenty percent of the market with that!” Jane makes a persuasive argument, and sells either her boss or the venture capitalists on the idea, and her team starts designing.
After about a year and a half, Jane’s team’s doohickey is ready to hit the market, so they ramp up orders for the chips they need. Unfortunately, unbeknownst to Jane, nineteen other companies saw the same opportunities that Jane did, at about the same time Jane did, and their design cycle was also about eighteen months. Thus twenty different teams all start ordering as if they are going to get 20% of the market, so demand skyrockets.
The teams, desperate for chips, start double- and triple-ordering chips in the hopes that that will increase their allotment from the semiconductor companies. (The fabs, however, have no way of knowing that the orders are bogus.) When Jane’s project does manage get chips, they hoard them in case the next shipment is delayed.
At some point, Jane’s management notices that Jane’s product has not, in fact, gotten 20% of the market, and that project is in fact losing money. Jane’s management sends in a hatchetman to clean things up. One of the first things that Ms. Hatchetperson sees is that Jane’s project is double- and triple-ordering product when the chips have become quite available and that there is a growing stockpile of those chips in the company warehouse. So Ms. Hatchetperson cuts all orders dramatically.
Meanwhile, the other nineteen projects come to similar conclusions at a similar time, so all twenty projects stop ordering at the same time, with devastating results for the fab. (This can happen astonishingly quickly: at my fab, we went from mandatory overtime in November 1984 to a two-week shutdown in December 1984.)
Conclusions
To summarize, there are four intertwined reasons why the semiconductor business goes through booms and busts:
- design cycles for products that use semiconductors are short
- it takes about the same amount of time for all companies to design a certain type of product
- semiconductor manufacturing yield is strongly negatively correlated with wafer starts
- it takes about as long for yield to come under control as it takes for companies to figure out that a product is a bust
Note that these were all true twenty years ago, when I was in the biz. It’s quite possible that yield management has gotten much better as a result of the computer revolution. (I certainly hope so!) But I doubt that the short design time, the similarity of schedules, and the optimism that a pet project will get 20% of the market has changed.