Machines From The Lunatic Fringe
By Philip Elmer-DeWitt
When Danny Hillis first appeared on the computer scene in the mid-1980s, it was easy to dismiss him -- and the odd-looking device he called the Connection Machine -- as part of the industry's lunatic fringe. The chipmunk-faced scientist from the Massachusetts Institute of Technology had achieved a certain local notoriety from tooling around the streets of Cambridge in a secondhand fire engine. As an undergraduate he invented a mechanical computer, made entirely out of Tinkertoys, that could play tick-tack-toe. And as a graduate student at MIT's famed Artificial Intelligence Laboratory, he spent much of his time worrying about things like how infants learn to recognize their mother's face.
Moreover, the concept behind the Connection Machine, a big black cube studded with red blinking lights, had the power and simplicity of an idea that is too good to be true. Most computers built over the previous 50 years had been designed to do one thing at a time; they funneled massive quantities of data through a single processor (the mathematical engine where the bulk of a computer's work takes place). Hillis proposed to break this computational logjam by replacing the single high-speed processor with large numbers of tiny computer chips that would attack the data in concert. The experts scoffed when Hillis argued that such "massively parallel" computers would soon move into the mainstream of computer science, surpassing in sheer speed and processing power even the famous supercomputers built by Cray Research.
The experts were wrong. Last week when Hillis introduced the Connection Machine's latest incarnation -- another sleek black box with red blinking lights -- most of his predictions had come true. Not only can the Connection Machine 5 lay claim to being the speediest computer in the world, having bettered the most powerful Crays on some problems by a factor of 100, but Hillis' company, Thinking Machines Corp., has become the leader in one of the industry's fastest-growing markets. The first seven customers for the CM-5, who paid from $1.5 million to as much as $25 million for models containing anywhere from 32 to 1,024 processors, include some of the world's premier computer users: the Sandia and Los Alamos National Laboratories; the Army High Performance Computing Research Center at the University of Minnesota; Syracuse University; the University of California, Berkeley; and the University of Wisconsin. Schlumberger, an oil-services company, ordered one to help interpret seismic data. American Express bought two for analyzing customer buying habits.
The success of the Connection Machine marks several milestones in computer science. One is the widespread acceptance of the parallel-processin g approach to computer design. "This was a watershed year for massive parallelism," says Gary Smaby, a supercomputer analyst at the Smaby Group in Minneapolis. There are more than half a dozen start-up companies selling parallel- processing computers of one sort or another. Both Digital Equipment and IBM, the two largest U.S. computer manufacturers, have endorsed the concept (IBM by forming a joint venture in September with Thinking Machines), and even Cray Research has begun work on a massively parallel supercomputer. Japan has selected the technology as the target for one of its long-term research undertakings, and at least three Japanese manufacturers -- NEC, Hitachi and Fujitsu -- are busy making their own Connection Machine-like computers.
Hillis' achievement also underscores the growing importance of supercomputers -- loosely defined as the most powerful number crunchers available at any given time. For years supercomputers were applied almost exclusively to national-security tasks, such as breaking codes or designing ever deadlier nuclear bombs. But the same computers that can locate a missile in outer space can also be used to find oil deposits in Prudhoe Bay, Alaska, and over the past decade a growing percentage of supercomputer sales have been to industry. Today supercomputers are used for everything from crash-testing cars to designing fuel-efficient aircraft.
The most eager consumers of supercomputer time, however, are scientists. Over the past five years, the number of researchers with access to supercomputers has grown almost a hundredfold, to more than 30,000, thanks to a network of supercomputer centers established by the National Science Foundation, the national laboratories and various state governments. In a wide variety of fields from astronomy to theoretical physics, computer simulation has replaced laboratory experimentation as a basic tool of scientific research. It is much easier to study the behavior of ionized gases in a computer simulation, for example, than it is to build a full-scale nuclear- fusion reactor. "We've whetted an awful lot of scientific appetites," says Larry Smarr, director of the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign.
But no sooner had scientists and engineers discovered the intellectual benefits of supercomputing than they found themselves bumping into the computational limits of the current machines. Everything they wanted to do, it seemed, required 1,000 times more computer power than the fastest machines could provide. Today's models, for example, are not able to determine the structure of a protein from a sequence of genes. They can map the earth's atmosphere or its ocean currents but not the interactions between the two. They can predict hurricanes, but not such smaller meteorological events as thunderstorms and tornadoes.
Last year President Bush's science adviser, D. Allan Bromley, compiled a list of 10 of these scientific problems, which he called "grand challenges," and asked Congress for more than $3 billion over the next five years to develop the computers and high-speed networks necessary to solve them. (The $638 million budgeted for 1992 is expected to be approved by Congress before Thanksgiving.) The centerpiece of Bromley's program is a research plan to build by 1996 a so-called teraflop machine, a computer capable of performing 1 trillion scientific calculations a second.
That goal may be reached sooner than anyone expected. The Connection Machine unveiled last week has a modular design that can be configured with anywhere from 32 to 16,000 processors. "We could build a teraflop machine today," boasts Hillis. In fact, a 16,000-processor CM-5 could deliver a peak speed of two teraflops -- if anyone could afford it. Using today's components at current prices, such a machine would fill a room the size of a small gymnasium and cost $200 million. Most analysts believe that the first teraflop machines will be purchased when their price drops below $50 million, sometime in the mid-1990s.
By then customers will have more than Thinking Machines to choose from. Intel, maker of the chips that run most IBM-compatible personal computers, is expected to announce its own teraflop initiative next month at a supercomputer convention in Albuquerque. Intel introduced a line of aggressively priced parallel supercomputers a year and a half ago and has nearly caught up to its Cambridge-based rival. One of its models, an experimental system called the Touchstone Delta, surpassed the top speed of the previous version of the Connection Machine last spring. Meanwhile, new massively parallel machines are expected over the next couple of years from Minneapolis-based Cray and such smaller companies as Kendall Square Research in Waltham, Mass., and Tera Computer in Seattle. By 1995, NEC, Fujitsu and Hitachi could be marketing their own teraflop machines.
Who actually sells the first teraflop computer is probably less important than who buys it. The big payoff from high-performance supercomputing -- both in profits and in international competitiveness -- will come when someone uses a Connection Machine, or a competing model, to design a wonder drug, a more efficient car or a cleaner-burning fuel. The new supercomputers are ready for delivery. It remains to be seen who will make the best use of them when they arrive.