Will We Keep Getting More Bits For The Buck?Otis Port
For three decades, the computer and semiconductor industries have thrived on the confounding economics of chipmaking. Every three years, the cost doubles for "wafer fabs," the ultraclean factories that fabricate chips on silicon wafers: It now stands at $1 billion. Yet the big investments have paid off by making possible still-tinier circuitry that further slashes the cost of computing. With chipmakers delivering four times the processing speed and memory capacity at roughly the same price every three years, the market has been gobbling silicon like teenagers scarfing potato chips.
All this might change, however, when the cost of a world-class fab climbs to $2 billion around the year 2000. Chipmakers won't be able to recover their investments unless they at least hold the line on prices--and prices might have to rise faster than performance. This would be a radical break in chip economics. Commodity chips such as dynamic random-access memories (DRAMs) "have always cost just a few dollars, regardless of their increased density," notes Craig R. Barrett, Intel Corp.'s chief operating officer. But when there's a $2 billion investment to recoup, "you can't sell them at $5 and make money."
Sticker prices have sometimes increased in the past, but the extra cost has always been more than offset by performance gains. In 1982, Intel's hottest microprocessor was the 286. It was rated at 1 million instructions per second (mips) and cost $360. Compare that with the Pentium, which Intel launched last year. It cost $950, but pumped out 100 mips. So the per-mips cost was only $9.50--a 97% plunge over 11 years.
To compensate for rising investment, chipmakers normally crank up output and spread the added expense over a larger number of chips. That's why the size of silicon wafers that end up getting sliced into dozens of chips has edged up steadily since the 1970s, from 2 inches to 6 and even 8 inches now--with 12 inches coming soon and 16-inch wafers due around 2000. These platter-size disks will hold four times as many chips as today's biggest wafers, which in the case of 16-megabit DRAMs is 212 chips per 8-inch wafer, according to Texas Instruments Inc. But even 16-inch wafers may not preserve traditional pricing trends.
GIGA-BETS. Chipmakers doubt that a modest increase in prices will dramatically dent chip demand, but that's not certain. And the uncertainty makes many producers queasy. Japan's chipmakers used to cough up capital even during recessions with no qualms. "Everyone pushed ahead and didn't worry what the systems guys would use it for," says Hisashi Hara, general manager of an advanced chip lab at Toshiba Corp., "because it was so obvious that they'd find a way to use it." But billion-buck sums seem tm push a hot button that wasn't tripped at $650 million, the previous tab for a fab. For the past two years, capital budgets in Japan were either frozen or trimmed--largely for lack of satisfactory answers to hard cost-justification questions.
Purse strings are now loosening somewhat, but the industry still isn't adding enough new capacity. Currently, there are 20 world-class fabs around the globe. Intel calculates that by decade's end, 22 more will be needed just to support growth in the PC business. Murray A. Goldman, senior vice-president of Motorola Inc.'s chip division, pegs the total need at twice that number. If they're right, shortages could compound any price hikes dictated by higher investments.
Put it all together, and cost containment is becoming chipmaking's No. 1 priority. Consultant G. Dan Hutcheson, president of VLSI Research Inc., figures just cutting corners could squeeze costs by perhaps 20%. For instance, instead of using gold piping to make sure no contaminants leach from the pipes into the ultrapure chemicals used to process wafers, Hutcheson argues that standard pipes with space-age filtering could do the job.
PREFAB FABS. Many other alternatives are being explored, with most chipmakers hunting for a magic recipe that would trim their investment increments by at least 30%. NEC Corp. is working on a convertible factory. It would consist of three modules, each producing a different chip generation. Every three years, the oldest module would be opened up and renovated for making the latest chips. Sematech Inc., the chip-industry consortium in Austin, Tex., and Semiconductor Equipment & Materials International, a trade association, are championing so-called cluster tools, a concept pioneered by equipment maker Applied Materials Inc. All production tools would have a standard connection that plugs into a central hub, where robots would shuffle wafers from one process to the next. Heretofore, every supplier has had its own interface, and integrating everything was a costly engineering task.
The most radical idea comes from TI. Working under a Defense Dept. contract, the Dallas chip giant has developed a $30 million setup that it says can match the productivity of a $1 billion fab. It's dubbed MMST, short for Microelectronics Manufacturing Science & Technology, and it reinvents the production of microchips almost from the ground up. For instance, instead of processing wafers in batches of two dmzen, MMST handles them one at a time--bringing "just-in-time" principles to wafer fabrication. "We have processed wafers in less than three days," slashing the usual cycle time by as much as 85%, boasts TI's Robert Doering, director of the MMST program. Toshiba's Hara expects the concept of single-wafer processing to spread, if only because 16-inch wafers will hold so many chips that only DRAMs and a few other high-volume chips can justify being batched.
If all the work to prune the cost of wafer fabs pays off--and especially if TI's MMST concept takes hold--the 21st century could open with a hectic period of small startups turning out innovative chips. But if that doesn't happen, America's small chipmakers will still be in demand, says Michael G. Borrus, co-director of the Berkeley Roundtable on the International Economy. In the coming age of one-chip products, he says, clever designs will be in greater demand than ever--especially if the cost of performance starts to increase.