Breaking A Computer Barrier
Three Ways to Speed Up Computers
By David E. Sanger
The New York Times
September 9, 1985
For two decades, computer scientists have periodically wrestled with a seemingly obvious technique for greatly speeding up computers: Rather than funnel a vast amount of data through a single processor, design a machine in which a host of processors attack a problem at once. The result, they knew, would be like installing a bank of toll booths across an interstate highway, instead of forcing hundreds of cars into a single lane.
But it never proved that easy. While it was possible to build multiple-processor computers that run a number of programs independently, no one seemed capable of focusing the power of several processors on a single task. They were constantly tripping over one another, lying idle for an eternity of milliseconds while one processor came up with the answer that the others depended on to complete the job.
Now, with a mix of new hardware and artful programming, a host of companies seem
on the threshold of breaking the barrier of parallel processing. The result, industry
experts say, is an emerging generation of computers 2 to 10 times faster than top-of-the-line
minicomputers, some approaching the power of small supercomputers. While no one
is predicting an immediate revolution, many believe that within five years, the
new technology could wreak havoc among the established order of computer makers.
'A Whole New Way of Thinking'
''We are on the verge of a whole new way of thinking about building computers,'' maintains Ronald H. Gruner, the 38-year-old chief executive of the Alliant Computer Systems Corporation, perhaps the most promising of a handful of parallel processing companies to spring up around Boston and Silicon Valley.
In an industry where technological revolutions are proclaimed weekly, Mr. Gruner's comments might be easily dismissed. But academics and other computer experts alike say that he may well be right and that the emergence of parallel processing could mark the biggest single change in computer architecture since the invention of the microprocessor 15 years ago.
Alliant's month-old eight processor FX/8 computer is only the latest, if perhaps the most elegant, of the new generation of parallel processors. But its allure lies in its compatibility with existing minicomputers, the workhorses used by engineers and scientists with the most pressing number-crunching problems.
Until now, most parallel processors, including a promising new entry by Intel Corporation, have required entirely new programming, a time-consuming task. But its first users say the Alliant machine is able to run hundreds of existing programs with little modification, mostly software designed for the Digital Equipment Corporation's powerful VAX minicomputers - sometimes at eight times a VAX's speed.
Whether the FX/8's capability will propel Alliant to the top of the computer heap on Boston's Route 128, as some predict, is far from certain. As makers of personal computers and mainframes alike have learned in recent years, the link between successful technologies and successful companies is tenuous at best.
What may spur the technology -and with it, a new sector of the computer industry - is the tremendous interest of top strategic planners in the Defense Department. Both the Defense Advanced Research Projects Agency and the Strategic Defense Initiative, or ''Star Wars,'' office are pouring millions of dollars into parallel-processing research. Any successful effort to track and destroy thousands of incoming Soviet missiles, they contend, could require systems of 1,000 to 10,000 processors.
Others are joining in, though often reluctantly. The International Business Machines Corporation says it is dubious that any major change in computer architecture is in the offing. But it has already built an experimental parallel processor in its research lab in Yorktown Heights, N.Y., and last month it gave $2.6 million to New York University to refine the design. On Tuesday, both the Encore Computer Corporation, a neighbor and competitor of Alliant in the Boston suburbs, and the Perkin-Elmer Corporation will reveal further details of systems they are about to market, though experts say their strength is the less complex field of ''multi-processing,'' handling a number of independent programs at once.
The enormously complex task of designing computers that can make the most efficient use of multiple processors at once is quickly absorbing some of the best talent in the industry. First among them is Mr. Gruner, Alliant's chief executive and a familiar figure to close readers of ''The Soul of a New Machine,'' Tracy Kidder's Pulitzer Prize-winning account of the birth of a Data General minicomputer in the late 1970's. A tall, soft-spoken Oklahoman who never took a computer science course, Mr. Gruner left college in 1969 to join Data General and quickly rose to become its top computer architect.
But it was Mr. Gruner's misfortune to lead Data General's ill-fated North Carolina team, whose daring mini-computer prototype was killed when it proved too late and too expensive for the company to market. The internal battle over the fate of that project, dramatic grist in Mr. Kidder's book, sparked a wave of departures at the company. By 1982, after a 13-year career at Data General, Mr. Gruner decided that he, too, was ready to strike out on his own.
He quickly drew in Craig J. Mundie, a software expert and member of the North
Carolina project. For six weeks the two parked themselves in the computer science
library at the Massachusetts Institute of Technology, poring over a decade's worth
of scientific papers.
'Germ of an Idea'
''We knew we wanted to start a company,'' Mr. Gruner said the other day, describing his searches through stacks of articles. ''We just needed the germ of an idea.''
The germ emerged from a series of papers written by David J. Kuck, a professor at the University of Illinois at Champaign-Urbana, and widely regarded as one of the world's leading experts on supercomputers. Soon after, they wooed the professor as a consultant.
Professor Kuck's pragmatic approach, a conviction that existing computer programs could be reorganized to lend themselves to attack by multiple processors simultaneously, proved irresistible to some of high-technology's leading venture capitalists. Hambrecht & Quist Inc. and Kleiner, Perkins, Caufield & Byers, both of San Francisco, and Venrock, the Rockefeller family's venture capital group, contributed $14.8 million in two rounds of financing.
Money in hand, Alliant's founders raided Digital. Prime Computer Inc. and Data General for top engineers, and immediately put them to work in a crash effort. Critical design decisions were made in late-night sessions at the tables of Acton's McDonald's and Burger King. When the first three prototypes emerged, they bore the names ''Whopper,'' ''Big Mac'' and ''McNugget.''
''Some computers are built on circuitry,'' Mr. Gruner grinned, ''and some are
built on junk food.''
Today the freewheeling air of a start-up venture still envelops Alliant's offices. The parking lot is never empty, the company provides free soda and snacks to the employees, and signs on the walls urge engineers to vent their competitive obsessions in an upcoming softball contest against a neighboring young company, described in the posters as ''wimpy personal computer types.''
There is nothing wimpy about the competition Alliant faces in the computer business. The pioneers in minicomputers, especially Digital, caught traditional mainframe manufacturers unawares, stealing a large share of the market and attracting a loyal base of users in corporate research labs and on university campuses. Few believe the feat can be repeated easily today.
Cray Research, the world's leading supercomputer maker, is working hard on parallel
systems. Sequent Computer Systems Inc. has already shipped 45 systems, primarily
used as ''multiprocessors'' that run several programs simultaneously. But C. Scott
Gibson, the Portland, Ore., company's chief operating officer, says about half of
the company's customers are using the Sequent machines as parallel processors.
Some Pioneers Have Troubles
A few pioneers have already run into troubles. Denelcor Inc., a Colorado concern that designed a promising parallel processor for the Government, is badly strapped for cash and considered a likely merger candidate. And while it is still among the promising contenders, Encore has suffered design setbacks and the departure of some of its talent even before its new ''Multimax'' computers, with 10 processors, have been shipped. Last month it lost a potentially large order from the Sperry Corporation because of product delays. Kenneth Fisher, the former president of Prime who now leads Encore, says the company ''is basically very, very strong.'' and is rapidly headed toward parallel processors.
Speed is at a premium. ''There is a window of opportunity now; the question is who will fill it before I.B.M. and Digital jump in,'' notes Jeffry Canin, a technology analyst for Hambrecht & Quist.
Alliant says it has already shipped six parallel-processor systems, costing from $270,000 to $1 million each, but refuses to make predictions about its revenues or earnings. Its products, however, have already attracted the attention of minicomputer users hungering for far more power.
At M.I.T., Thomas H. Jordan, a professor of geophysics, is awaiting delivery of his Alliant machine. A seismologist, he plans to use it on a project to distinguish between earthquakes and underground nuclear explosions, a task requiring tremendous number-crunching.
''I can't afford a Cray,'' he said. ''Having access to these kinds of machines is revolutionary. You have to think it will change the nature of computing.''
Leading supercomputer systems like Cray Research's disassemble a program and group similar calculations together. Thus, hundreds of addition problems will be solved first, then hundreds of multiplication problems, and the results are later reassembled. Theoretically, tremendously high speeds can be attained, but because most programs are not easily carved up, only a fraction of the machine's processing power is ordinarily used.
Computers run several programs at once, each on a different processor. The advantage is that users need not wait as long before their program reaches the front of the ''queue.'' Usually, however, the power of the processors cannot be harnessed together, meaning that multi-processors are of little help in speeding up extremely long calculations.
Machines like the new Alliant computer carve up a single program into various parts, parceling them out to different processors in the system. Thus, calculations are performed simultaneously, until one processor must stop and await results from another processor working on an earlier portion of the program. Again, only some types of programs are suitable, but tremendously high speeds can be achieved for relatively little cost.
Copyright 1985 The New York Times Company