The Next Great Leap In Computing SpeedRichard Brandt
Packaging technology doesn't normally stoke the passions of semiconductor engineers. The top guns in the business are the people who design circuits, not those who engineer the protective housings, or packages, for chips. But suddenly, packaging is winning the rapt attention of top scientists at Boeing, Digital Equipment, IBM, and Motorola, among others. For without big improvements in this obscure art, the perpetual increases in semiconductor performance that spur electronics-industry growth will slow drastically. "That's why packaging is going to be a hot topic in the 1990s," says George H. Heilmeier, chief technical officer at Texas Instruments Inc.
The secret to computing speed is size: The more compact the circuitry, the shorter the distance that electrical signals must travel--and the faster a chip processes data. Shrinking circuits have made desktop computers as fast as mainframes were just a few years ago. But now, a speed trap lurks around the bend. After racing through a chip's circuitry, electrical pulses get pumped through metal wires that tie a packaged chip to a circuit board. Then, they travel across the board to another chip. These journeys take mere billionths of a second, which didn't matter much before. But because tomorrow's desktops will be so fast, these delays could cut data processing gains by as much as 80%.
SNAGS AHEAD. The leading solution to this is multichip modules (MCMs). The idea is simple: Instead of putting one chip in each package, mount several--and eliminate those interchip delays. Along with the big electronics companies, a dozen or so startups are working on MCMs. Market researcher Dataquest Inc. predicts that multichips will bean $18 billion business by the mid-1990s--and that a third of all semiconductors will be housed in MCMs by the year 2000.
Before that happens, there are high hurdles to clear. Currently, chipmakers rarely sell unpackaged chips, since final quality checks can't be run until a circuit is in its housing. "We have exacting standards for quality," says Dana E. Krelle, marketing manager at Intel Corp. "We would not want that out of our hands." But there may be no choice: Few U. S. chipmakers produce all the chips likely to be used in an MCM.
There are other problems as well. If the quality level of the individual chips in an MCM is 97%, the chance of one defective chip showing up in a 20-chip module is almost 50-50. Since semiconductor packages generally can't be re-opened, a flawed module would mean throwing out 19 good chips. That's why MCM suppliers are working to develop innovative techniques for spotting defective chips before the module is sealed tight. Raychem Corp.'s Advanced Packaging Systems subsidiary, for example, attaches a tiny grid of metal wires to the raw chip. The wires are first used to test the circuit and then to attach it to the MCM.
In mainframes, the need for faster interchip communications became obvious a decade ago, and proprietary multichip modules now speed many big computers. Digital Equipment Corp.'s VAX 9000 mainframe has 13 MCMs, each with up to 76 chips. DEC says the modules double the 9000's performance. For its System/390, IBM uses modules holding up to 133 chips--and the MCM runs so fast and hot that it has to be refrigerated.
CRITICAL SPEED. Now, it's the turn of engineering workstations. MCM packages become crucial when a computer's heartbeat reaches 50 million pulses a second, or 50 megahertz--a mark that top-end workstations will soon hit. Motorola Inc. has a 50-Mhz microprocessor, and Intel has just unveiled a 100-Mhz design. A San Jose (Calif.) startup, nChip Inc., has a prototype five-chip module for SunMicrosystems Inc. workstations, and it runs 25% faster than the same chipsin ordinary one-chip packages. Multichip packaging, says David R. Ditzel, director of advanced development at Sun Microsystems, "is clearly coming."
Building a reliable workstation MCM that costs no more than an extra 20% to 30% will be one key. And besides quality concerns, there are technical ones, such as the most efficient way to attach the chips. IBM glues some upside down to its MCM. Newcomer Irvine Sensors Corp. stacks them and inserts them edge down.
Resolving such details is the goal of projects at Microelectronics & Computer Technology Corp., the Austin (Tex.) research consortium, and the Pentagon's Defense Advanced Research Projects Agency. Both are sponsoring multimillion-dollar runoffs between teams working on different approaches to manufacturing MCMs.
The joker in all this could be the Japanese. They, too, are perfecting multichip packages--and nurturing this market will be easier for Japan's vertically integrated suppliers, because they typically make a wide range of chips. By contrast, the fragmented U. S. electronics industry could get bogged down in turf battles over who lets go of naked chips and who does the packaging. Unless that's avoided, multichip modules could leave the U. S. a laggard in tomorrow's electronics markets.