Company Profile THE RISE and Fall THE 01 BRILLIANT START-UP THAT Some day we will build a think­ I~Z~~~~~ Thinking ing machine. It will be a truly NEVER GRASPED intelligent machine. One that can see and hear and speak. A THE BASICS Mach-Ines machine that will be proud of us. by Gary Taubes -From a Thinking Machines brochure

seven 'years a~ter. its The truth is very different. This is the simple proeessors, all of them completing In 19 90 founding, Thlllklllg story of how Thinking Machines got the a single instruction at the same time. To Machines was the market leader in paral­ jump on a hot new market-and then get more speed, more processors would lel , with sales of about screwed up, big time. be added. Eventually, so the theory went, $65 million. Not only was the company with enough processors (perhaps billions) protitable; it also, in the words of one IBM ntil W. Daniel Hillis came along, and the right software, a massively paral­ computer scientist, had cornered the mar­ Ucomputers more or less had been de­ lel computer might start acting vaguely . ket "on sex appeal in high-performance signed along the lines of ENIAC. Ifl that human. Whether it would take pride in its computing." Several giants in the com­ machine a single processor complete? in­ creators would remain to be seen. puter industry were seeking a merger or a structions one at a time, in sequence. "Se­ Hillis is what good scientists call a very partnership with the company. Wall Street quential" computers are good at adding bright guy-creative, imaginative, but not was sniffing around for an initial public long strings of numbers and at other feats quite a genius. He is also an inveterate offering. Even Hollywood was interested. of arithmetic. But they're seriously deli­ tinkerer, whose work has always been Steven Spielberg was so taken with cient at the kinds of pattern-recognition more fascinating than practical. On the Thinking Machines and its technology tasks that a two-week-old puppy can mas­ fifth floor of Boston's Computer Mu­ that he would soon cast the company's ter effortlessly-identifying faces or fig­ seum, for instance, is a minimalist com­ gleaming black in uring out where it is in a room. Puppies puter constructed of fishing line and the role of the in the film can do that because their brains-like 10,000 Tinkertoy parts. Hillis built it to Jurassic Park, even though the Michael those of all animals, including humans­ play and win at tic-tac-toe, which it in­ Crichton novel to which the movie was are "massively parallcl" computers. In­ variably does. His other work includes a otherwise faithful specified a . stead of looking at information one jig­ robot finger that can differentiate between In August of last year Thinking Ma­ saw-puzzle piece at a time, a brain a washer and a screw but is flummoxed chines filed for Chapter 11. It had gone processes millions, even billions, of by a piece of gum; a propeller-driven through three CEOs in two years and was pieces of data at once, allowing images jumpsuit that allows its wearer literally to losing money at a considerably faster rate and other patterns to leap out. walk on water; and a home robot con­ than it had ever made it. While a graduate student at MIT's Arti­ structed of paint cans, lightbulbs, and a What caused this high-flying company fIcial Intelligence (AI) Lab, Hillis, whom rotisserie motor. to come crashing to earth? The standard everyone knows as Danny, had conceived At the AI Lab, Hillis had become a explanation is that Thinking Machines of a for his thesis disciple of legendary AI guru Marvin was a great company victimized by the that would mimic that Minsky. The two were determined to sudden cutbacks in science funding process in silicon. Hillis called the device build a connection machine as a tool with brought about by the end of the cold war. a "connection machine": it had 64,000 which to develop software programs for

1995, No.3 Inc. Tcchnology 61 . Because the cost others eventually agreed to kick in a total he quit. Emotional decision making would be prohibitive for a university lab­ of $16 mi Ilion to the venture. would last almost until the company fell. oratory, they decided to form a company. In May 1983, despite the lack of a busi­ They went looking for help and found ness plan, the company was founded and n the first few years it didn't seem to Sheryl Handler. took up shop in a dilapidated mansion I matter. Thinking Machines didn't need Handler had participated in the start-up outside Boston that once was owned by to make good business decisions because of the Genetics Institute, a Harvard-based Thomas Paine, the author of the Revolu­ it had the Defense Advanced Research genetic-engineering firm. Her back­ tionary War pamphlet Common Sense. Projects Agency. A research arm of the ground was eclectie: she haa studied inte- Hillis and Handler called their new Defense Department, DARPA was looking company Thinking for computer architectures that would en­ Machines because, able tanks, missiles, and other weapons to says Hillis, "we recognize enemy targets and understand wanted a dream spoken orders. In 1984 Hillis and his col­ we weren't going leagues at Thinking Machines repackaged to outgrow." As it Hillis's thesis and pitched it to DARPA. turned out, there The agency responded by offering the was never much company a multiyear $4.5-million con­ danger of that. tract. Now all Thinking Machines had to The new com- do was build one of the world's fastest pany's managers computers in two years' time. immediately got The company promptly went on a hir­ into a disagree­ ing binge. lts prime hunting grounds were ment over the mar­ the computer-science departments of MIT, Brain man: wanted to build machines that could ket for supercom­ Carnegie-Mellon, Yale, and Stanford­ achieve emotions-not machines that would solve companies' puters. Hillis and which happened to house four of the problems. Handler (Minsky world's leading AI labs. Everyone, from quickly became a programmers to administrative assistants, rior design, held a master's degree in figurehead at the company) wanted to de­ had to be interviewed by Handler, who landscape architecture from Harvard, and sign a machine strictly along the lines of had a very specific, if mysterious, idea of at the time was pursuing a doctorate in Hillis's thesis, a machine that would have who would be good enough to work for city planning at MIT. She was also run­ its maximum impact as a research tool for Thinking Machines. (Many researchers ning her own nonprofit consulting firm, scientists studying artificial intelligence. later reported that once they were hired, specializing in third-world resource plan­ (Hillis envisioned his machine eventually they never got to speak to Handler ning. She had a taste for classical music becoming a sort of public-intelligence again-even when they were alone with and a fine appreciation for style. She'd utility into which people would tap their her in an elevator.) even been the subject of a Dewars Profile home pes, thereby bringing artificial in­ In fact, Thinking Machines was be­ that ran with the quote "My feminine in­ telligence to the world.) Howard Res­ coming Handler's aesthetic creation as stinct to shelter and nurture contributes to nikov, a research director recruited by much as the Connection Machine was my professional perspective." Minsky, on the other hand, argued for a Hillis's. In the summer of 1984 the com­ Handler also had a talent for cultivat­ more f1exible architecture that could sup­ pany moved into its new home-the top ing friendships with brilliant and famous port whatever style of computing was two nOOfS of the old Carter Ink Building people. One of her Genetics Institute col­ needed to solve real-world problems. Af­ in Cambridge, Mass., a few blocks from leagues later called her a "professional ter all, the more problems the machine MIT. Handler personally oversaw the de­ schmoozer." She quickly proved her use­ could solve, the more sales prospects sign of the office space, insisting that each fulness by connecting the people who there would be. office be painted a different and. specific would build the Connection Machine For a year, while the argument went on, color. Huge open spaces were created to with CBS founder William Paley. Hillis, the company did nothing. Finally, Han­ stimulate idea sharing and creativity. A t-.linsky. and Handler pitched the idea to dIer and Hillis won out. "We had all sorts plush cafeteria was put in, complete with "aley and CBS president Fred Stanton in a of reasoned discussions," says Resnikov. a gourmet chef. Couches were scattered meeting. to which Hillis wore his custom­ "and then emotional decisions were fun­ throughout the offices so that researchers ary jC4II1\ illll! T-shiI1. Still, he managed to damentally made by Sheryl and Danny." could take naps or even sleep there imprcs, the tdcvisioll moguls, who with Resnikov lasted another two years before overnight, which many of them did. And

62 Inc. Tcchnology 1995. NO.3 the soft-drink machine was wired to a ter­ weight outside AI research. Unfortu­ predicting earthquakes, revealing the nu­ minal. Researchers who wanted a drink nately, according to Resnikov, the deci­ ances of quantum mechanics. The prob­ simply typed in their choice. sion to tailor the CM-I to the AI "nonmar­ lems didn't require artificial intelligence, In short, Thinking Machines was be­ ket" cost Thinking Machines three years just enormous computing power. coming a hacker's paradise. The thinking, in the real-world marketplace. The official name of the new project says Lew Tucker, one of the company's In April 1986, Thinking Machines an­ was the High Performance Computing research directors, was that "if they were nounced the arrival of the CM-2, a ma­ and Communication (HPCC) program, fed, they'd practically live at Thinking chine the scientific community actually and DARPA was the lead agency, with a Machines." If Hillis disapproved, he could use. The CM-2 was able to nm FOR­ projected budget of several billion dollars didn't make it known. Having taken to TRAN and to do floating-point operations. through 1996 to accomplish its goals. At commuting in an antique fire engine, he It was also a piece of work artistically: a could hardly play the pragmatist to Han­ five-foot cube of cubes-done up in what dIer's stylist. Thinking Machines employees called In May 1985, Thinking Machines an­ "Darth Vader black"-in whose innards nounced the impending completion of the red lights flickered mysteriously. But the first Connection Machine, the CM-l. The machine's exotic massively parallel tech­ announcement would be made on the nology still needed special software, third floor of the Carter Ink Building. which meant its users had to Jearn new Handler had every surface on the new programming techniques. The CM-2 floor repainted a slightly di fferent shade might be more like the human brain than of mauve. When it was done, she wasn't a sequential computer like the Cray was, satisfied. So she had her researchers but scientists knew how to write programs and scientists for the Cray. Many of Thinking Ma­ paint it again. chines' first customers, says Dave Waltz, The CM-l who ran the company's Al group, did was an Al most of their computing on the lloating­ researcher's point processors, ignoring the 64,000 sin­ dream. Un­ gle-bit processors. ¥ fortunately, As a result, there still wasn't much of a few AI labs market for Connection Machines. But'• could afford thanks to the support of DARPA, which Dream machine: The CM-1 a $5-million continued to broker deals, Thinking Ma­ contains thousands of tiny computer, chines didn't have to seriously contem­ processors instead of one and, as Res­ plate building a machine that had a nat­ big one. nikov had pre­ ural market. "Our charter," says Tucker, dicted, hardly "wasn't to look at a machine and figure anyone else was interested. When it came out tile commercial profit. Our charter to general scientific computing, the CM-l was to build an interesting machine." But was "a dog," in the words of Gordon Bell, the definition of interesting would soon a computer guru and architect of the fa­ change. mous VAX computer at Digital Equipment Corp. It had no facility for lUnning FOR­ n the late 1980s, DARPA and the Bush TRAN, the de facto standard computer lan­ I administration, having accepted the fact guage of science; nor could it do what are that the end of the cold war had reduced Connection machine: CEO Sheryl Han­ known as "floating-point operations," the the urgency for military supercomputing, dier had lots of contacts. But could she operations that manipulate numbers in sci­ came up with a new challenge for parallel run a company? entific computation. computing. They began to talk about Thinking Machines sold seven CM-Is, solving what D. Allan Bromley, the presi­ but only because DARPA broke red and dent's science adviser, dubbed "grand the top of the list: building a computer ca­ subsidized most of the deals. If the com­ challenge" scientific problems: modeling pable of a teraflop-a trillion floating­ pany was going to stay in business, it the global climate, analyzing the folding point operations per second. would need a machine that could pull its of proteins, mapping the human genome, Not surprisingly, Thinking Machines

1995. No.3 Inc. "lcclmology 63 'Vendors had an inside track on getting a chunk of handed money business supercomputer group, an idea the projected budget. While other com­ by the that appears at first to be a no-brainer. But puter companies were out wooing cus­ at Thinking Machines the idea got stuck tomers, Handler had been cultivating a government in endless discussions. Hillis and Handler friendship with Bromley. As soon as have no interest already were bitter about having to target Thinking Machines promised it would general scientific computing rather than have a scaled-down version of a teranop in solving artificial intelligence; they weren't about machine ready by 1992, the agency customers' to jump on the idea of servicing mere awarded the company an initial contract merchants. Hillis later complained about of$12 million. problems,' the injustice of a world where "the real In the meantime, several computer growled KSR's money is in handling Wal-Mart's inven­ companies were exploring a new technol­ tory rather than searching for the origins ogy-a compromise between the comfort Burkhardt. of the universe." of sequential computing and the perfor­ Nonetheless, thanks to DARPA, Think­ mance of massively parallel machines. A ing Machines went into the black for the sort of "moderately parallel" design, tl;le first time. In 1989 the company reported a technology entailed stringing togethe~ a profit of $700,000 on revenues of $45 smaller number of the powerful, cheap' plication" for parallel computers. With the million. Handler promptly signed a 10- off-the-shelf used in pes country in a recession, businesses needed year lease with the Carter Ink Building for and workstations-rather than the thou­ every competitive advantage they could a whopping $6 million a year-about $37 sands of highly customized but less pow­ get, which meant knowing their cus­ a square foot. (Lotus Development Corp., erful processors used in the Connection tomers' preferences and buying habits in which was virtually across the street from Machines-into a single supercomputer intimate detail. They had begun to collect Thinking Machines, was paying $8 a that would work with existing software. all conceivable data and were feeding square foot.) Thinking Machines also The cost advantages of using off-the­ them into their mainframes, looking for hired another 120 employees, bringing shelf chips, as well as the functional ad­ any insight that would help them maxi­ the total to over 400. Meanwhile, the vantage of running existing software, mize profits. But it sometimes took main­ company had developed an image as one seemed overwhelming-especially con­ frames hours, even days, to churn out the of the leading high-tech companies in the sidering the fact that few customers out­ answer to a single question. So large com­ country. It was, says Stephen Wolfram, side the tiny AI community had much in­ panies were beginning to check out paral­ who founded the highly successful soft­ terest in Thinking Machines' massively lel computers. ware company Mathematica, "the place parallel design. Even Hillis that foreign trade delegations would come eventually came around and to visit to see where American business chose the moderately parallel was at these days." design for the company's next Yet competition was looming. Cray generation of machine. Unfor­ Research launched a crash program in tunately, the old dream died 1990 to get a massively parallel machine hard: the decision came only af­ on the market within two years. IBM was ter 18 months of internal bick­ doing the same. Even Fujitsu Limited, ering. Once again, the company one of Japan's major supercomputer was off to a late start. manufacturers, was in the process of What's more, there were opening a parallel-computing lab, look­ signs that the company was still ing toward marketing a 1,000-processor chasing the wrong market. In­ Outvoted: Howard Resnikov, a research director machine. dustry analysts in 1992 were recruited by Minsky, argued for a more flexible ar­ If there was ever a time that Thinking projecting that the growth in chitecture that could support whatever style of com­ Machines could, and needed to, put itself ~ supercomputers was not in sci­ puting was needed to solve real-world problems. on a solid financial and competitive foun- ~ ence but in business applica­ dation by merging with a deep-pocketed ~ o tions-in particular in what's known as In fact, Thinking Machines had sold company or by going public, it was now. ~ "database mining," an area that could two Connection Machines to American But Handler nixed all deal making. She (5 well become, as IBM parallel-computing Express. That got management at Think­ felt the company could get a wildly suc- ~ expert Art Williams put it, "the killer ap- ing Machines talking about starting a cessful teranop machine out on its own. ~ 64 Inc. Techno)o/!y 1995, No.3 s the company forged ahead with its no interest in solving customers' prob­ mained at its helm, he engineered her A frantic effort to bring the new ma­ lems," he growled. ouster. chine out on time, the eorporate culture An embarrassed Bush administration Fishman focused the company on the started to shift from openness to paranoia. put an end to Thinking Machines' DARPA business market and began looking for a Employees weren't allowed to diseuss the gravy train. For the first time the com­ partner. Sun and IBM were interested, machine with one another in the cafeteria. pany had to sell its machines on their mer­ says Tucker, but weren't willing to take Customers were kept in the dark. The new its in an open market. At the end of 1992, on Thinking Machines' mounting deht, maehine was dubbed the CM-S, to foil Thinking Machines reported a loss for the which included six more years of rent at hackers acting as corporate spies who pre­ year of $17 million. The CM-S wasn't the Carter Ink Building, a $36-million sumably would be rummaging through selling, and the company was hemorrhag­ commitment. the company's files looking for a nonexis­ ing money. Hillis was no longer spending In mid-August, Thinking Machines tent CM-3. much time in the office. The first round of filed for bankruptcy protection, and Fish­ Thinking Machines announced the layoffs had started. Salaries were frozen. man resigned. Soon Hillis himself left the CM-S in October 1991. Hillis claimed it Requests for new laptop computers were company that had been founded around had the highest "theoretical" peak perfor­ being denied. his thesis. Thinking Machines would mance of any supercomputer ever, if you Meanwhile, Handler had an enormous rcemerge as a small software firm selling added enough processors to it. The real­ marble archway installed in the atriulll of programs for its former competitors' par­ ity: at the time completion of the CM-S the Carter Ink Building. When a national allel computers. was announced, the machine was slower supereomputer conference was held in As late as 1989, says Fishman, Think­ than i'ls predecessor, the CM-2. Among Seattle, she decided to stay in San fran­ ing Machines was still three years ahead other problems, the standard chips the cisco and commute to Seattle from the of the rest of the world in parallel­ company had chosen weren't ready, so swank Stanford Court Hotel. She com­ processing technology. "Whi Ie others some machines had to ship with slower, missioned a $40,000 logo design for a caught up," he says, "Thinking Machines earlier-generation chips. Meanwhile, CM-S sweatshirt and then rejected it. was losing time, losing customers, and competitors like , Kendall Square While the eompany was sinking, she fo­ not moving on to the next generation." Research (KSR), MasPar Computer, and cused her attention on putting out a cook­ Had the CM-S been built without the mis­ nCube were starting to ship faster super­ book with recipes from the company~s cues and the wasted time, the company computers. More than ever, Thinking Ma­ now-infamous cafeteria. Increasingly might have gone on to live up to its con­ chines was depending on its DARPA edge paranoid, she had a video camera aimed siderable promise. But, as one of the com­ to move its products. at her personal parking spot and, by sorlc pany's senior scientists would later put it, Then, in August 1991, as DARPA was accounts, made people take meetings with what if pigs could fly? 0 about to start the process of determining her in her parked car. She hired a body­ which supercomputer vendors would win guard, telling her colleagues that she had Gary Taube.\' is a New York-based science the lion's share of its planned spending recei ved death threats. and technology writel: His IIU1.1't recellt spree, the Wall Street Journal broke the Some members of Thinking Machines' book is Bad Science: The Short Life and story that the agency had been playing fa­ board suddenly seemed to realize that the Weird Times of Cold Fusion vorites. It turned out that DARPA had sub­ person who had been running the sidized-sometimes to the tune of the en­ company all those years had no tire purchase price-the sale of somc 24 business skills. Thc board dis­ Connection Machines in recent years. cussed dumping Handler, but shc The subsidies added up to a gi ft to managed to get her biggest ene­ Thinking Machines of $SS million-20% mies there kicked oiT. of the company's lifetime revenues to that In early 1993 a new president point. was brought in, but Handler, who DARPA had greased Intel's supercom­ remained CEO, quickly got rid of puting wheels too but had left the rest him. Later in the ycar a lawyer of the supercomputer industry to fend for named Richard Fishman was itself. And now the other players were hired as president. Fishman was hOWling. Perhaps the clearest and most a longtime friend of Handler, damning criticism came from KSR but when he realized that no out­ founder Henry Burkhardt: "Vendors sider would fund the sinking Swan song: Without DARPA's push, the CM-5 handed money by the government have company while Handler re- didn't sell.

1995, No.3 Inc. 'Icchnology 65 L Y I

Gordon Bell claims to be the winner of the . famous bet on supercomputers that he made B with Danny Hillis five years ago. But the bigger issue in the dying

supercomputer

market is not

who has the

B fastest machine, but whether anyone will still y

w be making Big Iron in the future.

E

s A S THE BEL L tolls at the end of to die as a supercomputer maker 1995, it tolls for Danny Hillis. H and reSUlTect itself as a software com­ Hillis has been on one long, A pany. The twin forced march of Hillis strange trip in the five years since T and TMC symbolizes the seismic shift Z Gordon Bell bet him a crow-eating that has rocked the traditional super­ essay about the future of supercomput­ computing market. ing. Hillis, the founder and fanner Bell didn't exactly predict that chief scientist of parallel processing shift, but he did bet on Hillis' comeup­ pioneer Thinking Machines Corp. pance. Bell, the legendary computer (TMC), Bedford, Mass., no longer holds designer who created the PDP and that position. He has returned to his V AX series of machines at Digital roots at the Massachusetts Institute of Equipment Corp., Maynard, Mass., and Technology in Cambridge, Mass., this who is now dispensing wisdom for time at the younger and more hip Microsoft Corp. from his office in Los Media Lab rather than the Artificial Altos, Calif., bet Hillis that massively Intelligence Lab, dominated by Marvin parallel supercomputers might not Minsky, where Hillis first cobbled actually be the greatest revolution to together a design for a massively paral­ hit the scene since the invention of lel computer. rock and roll. Specifically, in 1990, Bell TMC itself is just emerging from bet Hillis that in the last quarter of the dark side of a Chapter 11 bank­ 1995, the highest number of sustained ruptcy pleading. But to do so it had MFLOPS (millions of floating point

JANUARY 1996 UPS IDE 39 operations per second) would be generated by a last year or so, three more Big Iron companies joined machine with fewer than 100 processors, rather the list of roadkill: TMC (at least as a Big Iron com­ than by a machine with many processors (more pany), and Cray Computer. than 1,000). The wager concerns only super­ The high-end market, in fact, never ended up as computers, or Big Iron machines costing more than anything more than a tiny blip on serious financial $1 million and used for scientific purposes. That bet radar screens. The Smaby Group, Minneapolis, esti­ was chronicled in the January 1992 issue of UPSIDE. mates last year's entire market for high-perfor­ The judge, jury and keepers of the bet are John COMPUTERS mance (scientific/technical/engineering) computing Hennessey, a professor of electrical engineering and 1985-1986 at $2.05 billion. The company projects that in 1999 at , Stanford, the market will barely creep over $3 billion. But the Calif., and David Patterson, a professor and chair of true, high-end supercomputer segment of that mar­ the computer science division at the University of ket is beginning to crumble away. Smaby Group California, Berkeley. The loser has to 'fess up his projects that the top end of that market-machines humbling defeat to the world in writing. costing more than $5 million apiece-will decline The prohibitive favorite when the bet was by 6 percent in that time. Similarly, Chris Willard, made, Hillis is now such an underdog that Las manager, high-performance technology, in the Vegas bookies wouldn't touch him. "The only rea­ Mountain View, Calif., office of Framingham, son I might lose is if people get a few huge machines Mass.-based Intcrnational Data Corp., predicts that only doing floating point operations and lock them by 1999 the revenue for the traditional supercom­ in a room and they just sit there and grind out puter market will drop to $767 million from the MFLOPS,/I says Bell. "They'd be grinding out a shit­ 1994 total of $877 million, a negative 2.6 percent load of numbers just to win the bet./I A Cray T3E is SYSTEMS · compound annual growth rate. still probably the most powerful machine not on 1983-1989. Market watchers, however, are not yet ready to wheels. liThe [massively parallel] stuff is a bust,/I bury the entire Big Iron market. Willard concurs says Bell. "There's no market for it because no one that the high-end market "is not really dead, but can get applications for the machines. /I it's certainly not growing./I If Bell, whose volume and frequency of pontifi­ But if it's not a dinosaur, it's at least a white cation make the late Howard Cos ell seem mute by rhinoceros at the top of the Endangered Species

II FRO M T HAT CHAN GE D MAR KE T STAN DPOI NT, I TH INK WEB 0 T H LOST," SAYS HILLIS. uTHE BET MAY FALL IN THE DEAD ZONE."

comparison, had said that three years ago, even his List. That's a huge contrast from the days when most ardent admirers would have questioned his Hillis and Bell were betting on which type of super­ sanity. Until recently, many people considered computer would get bragging rights to the title of massively parallel machines to be the savior of fastest in the world. After all the hype and promise, high-performance computing. most of the racehorses collapsed before reaching But the trip to the future turned out to be even POINT the finish line, and Hillis. and Bell have found stranger than Bell might have predicted. It's not just themselves betting over tombstones. "From that the massively parallel machines that have disap­ SYSTEMS changed market standpoint, I think we both lost,/I pointed their creators and investors. We're talking 1970-1991· says Hillis. "The bet may fall in the dead zone./I upheavals in the high-performance computing world-alias Big Iron-that no Richter scale could THE BIGGER THEY ARE ... measure. And the aftershocks nave only just begun. It was one hell of a ride, though. A few short years "There is no future for the Big Iron systems, /I ago, supercomputers and were declares Michael Burwen, director of the Palo Alto as hot as the World Wide Web is today. Venture Management Group, a high-performance comput­ capitalists were pouring in money and new compa­ ing market research company in Palo Alto. nies were sprouting up every few weeks. Now even Hillis is humbled. "The big surprise But today, Cray Research Inc., Eagan, Minn., is [that] the supercomputing market basically doesn't still towers over the traditional supercomputing exist anymore as a definable market, /I he said last business. And perhaps it should. Cray Rese~rch has October at the Media Lab's 10th anniversary party. defined the excitement over supercomputing since "It's very clear that the dinosaurs are dying./I it was founded in 1972. As soon as the first Big Iron Many of the carcasses, in fact, already line the box sprang from Seymour Cray's head, it was crys­ Information Highway: Alliant, Scientific Compu­ tal clear that the thing had attitude-simultane­ ting Systems, Multiflow, Floating Point Systems ously mysterious, ethereal and fascinating. There and Supercomputing Systems Inc. And in just the was a unique majesty to the word supercomputer.

40 UPS IDE JANUARY 1996 It crunched numbers and solved prob­ $40 MILLION VAPORWA,RE, lems that previously had been the stuff of dreams. ANY 0 NEt How's this for a hot prospect? The genius of supercomputers, Seymour Cray, changed the world-and "The Company is a development stage Tera is also negotiating a contract with enterprise that had an accumulated loss ARPA to jointly fund the development of made one hell of a lot of money doing of approximately $9.6 million as of June certain components of its next-genera­ it-as the master of the bipolar-logic 30,1995. The Company has experienced tion MT A system. superprocessor. When the supercomput­ net losses in each year of operations and In addition, the San Diego Super­ ing universe centered around single-pro­ expects to incur substantial further loss­ computer Center (SDSC) has submitted cessor systems, he who made the fastest es while it builds its MTA (Multithreaded a proposal to ARPA to purchase Tera's Architecture System) system prototype first production MTA system. ARPA has CPU ruled. In that domain, Seymour and commences production, and possibly told Tera that it plans to exercise an Cray was unbeatable. thereafter. The Company has had no rev­ option under the January 1995 contract But in recent years, even the mighty enue or earnings and does not expect to to buy an MT A system to place at SDSC. Seymour has begun to look outmoded. recognize revenue from the sale of its (Calls to ARPA officials seeking comment While most people recognized that MT A system sooner than the second half about this expenditure of taxpayers' of 1996, if ever." funds were not returned.) killer microprocessors are taking over Sounds like the kind of statement This kind of backing indicates the field, Cray refused to yield. He tena­ that just builds confidence, doesn't it? there's a new guru in town. "This is cious]y-some say stubbomly-clung to Well, that's the wording from the absolutely startling," says Bob Stern, a his mission of making the fastest, most September 25 IPO statement of Seattle­ Washington-based IT consultant. "There powerful superprocessor. In 1989, the based Tera Computer Co. The company seems to be a serious national invest­ went public in order to raise money for ment in Burt Smith [Tera's founder, pioneer left the company he founded to its nonexistent supercomputer, whose chairman and chief scientist]. Seymour start another venture, Cray Computer selling price it expects to set between Cray was in the same position six Corp., to create even more powerful $5 million and $40 million. months ago-he needed $25 million but superprocessors from gallium arsenide. Since its December 1987 inception, couldn't get it." Bu t the superprocessor crusade, Tera has spent $27 million to develop Stern also notes that this seems to the MTA system. More than $18 million be government supercomputer business which had worked so well for so long, of that was a gift from the Advanced as usual. "This obviously shows that the had become an anachronism. In 1990, Research Projects Agency (ARPA). Still, days of the 'state computer' [Gordon Lawrence Livermore Laboratory agreed the prospectus notes that the MT A sys­ Bell's term for companies kept alive by to pay $42 million to be the proving tem "has been subject only to computer the federal government] aren't dead. It's ground for an eight-processor Cray-2 simulation and the Company has not yet amazing that Smith pulled this off." built its initial prototype." Get real, counters Patrick W. and Seymour's newest creation, a 16- So who would buy stock in such a Grady, senior vice president in the San processor Cray-3. But the Cray-3 was a company? Tera put 850,000 units up for Francisco office of Rochester, N.Y.­ stillbom prototype. sale, each consisting of two shares and based H.J. Meyers & Co. Inc., the offer­ Because of Seymour's reputation, one warrant. The stock price was $12 ing's underwriter. Grady claims that he the mere possibility of completing the per unit. The warrant entitles the holder has been approached by many of Tera's to purchase, at any time over a five-year competitors several times and hasn't Cray-4 kept the company alive long period starting September 25,1995, one given them a damn thing. "I'm com­ after less tolerant-or hero-worship­ share of common stock at $7.20 per pletely unconvinced that Cray Research ping-creditors would have ripped out share through March 24, 1998, and at holds the future of high-performance its life support systems. Death came last $8.40 per share through September 24, computing," he adds. April, not with a bang but a whimper 2000, when the warrants expire. Tera's management insists it does A lot of people bought. The offer­ have the answer. According to Vice when financing finally dried up. ing was a monster success, raking in President Gerald Loe, customers are Steve Chen, another supercomputer $8.55 million ($9.9 million if the under­ desperately seeking general purpose genius from Cray Research,' also ran writer's over-allotment is exercised). scalable parallel machines with large­ away to fight another day. Reveling in "We didn't have much difficulty scale memories. Tera promises to create his X-MP and Y-MP glory, he formed selling it because we've got a very excit­ such machines because it claims it can ing and contrarian story to tell," says Jim solve the memory latency problem that Supercomputer Systems Inc. (SSI) after Rottsolk, Tera's president, CEO and slows down other architectures. Grady convincing IBM that he had the cure for CFO. "It's unusual to go public when contends that cure extends all the way its supercomputing sickness. After four you have no revenue. But it's the pro­ to the desktop. years amI possibly as much as $250 mil­ mise of the product that interests "We know the machine's not lion from Big Blue, however, Chen investors and the government." built," Rottsolk concedes. "We know no Indeed, the government seems one knows if the hardware works. We proved incapable of walking the walk. especially interested. Aside from helping know that everyone could lose all their He allegedly produced a prototype based to fund the company's research, ARPA is money. But we strongly believe that on superprocessors, but he never sold a first in line to buy its products-once we're going to release our prototype in machine. So a sadder but wiser IBM, they're done. Last January the agency the first quarter of 1996 and deliver it now a self-made supercomputing heavy­ signed a contract with Tera that pro­ by June 30. vides ARPA with options to purchase up "Right now, though, you've got to weight, pulled the plug in 1992. to $20 million of MTA systems for early take it on faith because the product Unlike his mentor, however, Chen evaluation over the next three years. doesn't exist." Amen.-W.S. recently demonstrated that he may not

42 UPS IDE JANUARY 1996 ~ti1l be crazy after all these years. Funded entirely their big brothers. Convex Computer Corp. was by MCSB Systems PTE Ltd., a Singapore-based started in 1982 as a venture capital-financed mini­ technology conglomerate, Chen Systems Corp. (for­ supercomputer alternative to the expensive, mas­ merly SuperComputers International) in Eau sive Big Iron boxes that dominated the high-perfor­ Claire, Wis., began beta testing its Pentium Pro­ mance computing market. The plan was to build based Chen 1000 server line last April. Its eight-pro­ 1/ affordable" supercomputers-in other words, cessor machine was released on September 18, smaller, cheaper and more efficient machines. mostly to yawns. The company at the time claimed Everything proceeded smoothly for almost a it had 20 orders for the new machine, and if so, decade. Convex went public in 1986 and its stock that's an impressive debut. But it's hardly a super­ traded at more than $20 by 1990. It had an im­ computer. As one source who declined to be identi­ pressive string of consecutively profitable quarters. fied asked, "What's the big deal about another Intel The company was all the rage on Wall Street. -based machine?" But when the Cold War went down, it also Unable to match Cray at the high end, most brought Convex' with it. The Defense potential rivals went low. They gradually realized Department stopped writing checks as freely as that more microprocessors meant more power to it did in 1983, when defense official Richard the people. They coupled a few, then tens, then Perle told a Congressional hearing that an Apple dozens, then hundreds in a single system. They II was capable of starting World War III. Convex's didn't fare any better, however. profits turned to 10sses-$140 million since No company rose higher quicker and fell lower 1993, including $15.3 million in the first half of faster ~han the classic of parallel processing, this year. Money went out faster than it came in; Thinking Machines. TMC at the beginning of the revenue plummeted to $144.2 million last year decade had sales of $65 million and thoroughly from $231.8 million in 1992. "I hate to use the dominated the burgeoning parallel processing mar­ term, but this paradigm shift is changing every­ ket it had created. The future seemed limitless. thing that has been the norm for the last 10 Hillis had his own worshippers, including years," says Steve Wallach, co-founder and senior Steven Squires, director of the Computer Systems vice president of technology at Richardson, Tex.- WHAT HAPPENED TOT H E MARKET THE REAL PROBLEM WA S THAT THE MARKET NEVER REALLY HAPPENED.

Technology Office at the Advanced Research based Convex, now the Convex Technology Projects Ageney (ARPA). Through targeted grants Center of Hewlett-Packard Co., which bought and sweetheart deals, ARPA came very close to vio­ the company last September for $150 million. lating the spirit, if not the letter, of government contracting regulations, enabling Thinking BACK FROM THE DEAD Machines to lead its channed life. What happened to the market? The real problem Since its May 1983 founding on little more was that the market never really happened. The than a wing and a prayer, TMC never had to worry breathless anticipation of the military turned into a about business plans, competitive strategies or the panting enthusiasm from entrepreneurs and ven­ color of its balance sheet. Every time the company ture capitalists, and none of it was really deserved. leaked, ARPA patched the hole. Thinking Machines and others discovered that cor­ But the days of the future passed in a nanosec­ porations are not as free-spending as the military. ond. With the Big Red Menace broken and no Says Bell, "TMC got in so much trouble because in longer the justification for an unlimited military the beginning they extrapolated, from a few sales budget, ARPA pulled the plug. directly related to government placements, that The company's Connection Machine couldn't there was a market for huge processing machines. cut it in the real world. TMC was also wracked by There never was a real market there." management turmoil. In August 1994 it sought For a while, the military enthusiasm seemed shelter in the Chapter 11 bankruptcy womb. infectious. There was a certain amount of prestige Kendall Square Research, another one-time to owning a supercomputer. Apple Computer Inc. high-flier in the massively parallel world, even even bought a Cray for its research efforts. But as drew Bell as investor and consultant. Founder the chill of the Cold War thawed, million-dollar Henry Burkhardt was a terrific technologist and a supercomputers became as popular as $500 toilet terrible accountant. After revealing accounting seats. Business process reengineering, a buzzword irrcgularities in 1995, it also filed for Chapter 11. that mostly means cutting expenses and people, The fate of the minisupers paralleled that of accelerated the trend. The CIOs and the MISers

JANUARY 1996 UPS IDE 43 began buying solution cycles, not just MFLOPS. other groups and became the Business Systems Users stopped genuflecting at the high-end altar Division. That entity's sole offering is the CS6400 and started asking what the machines could actual­ enterprise database server, a multiprocessor system ly do for them. based on the RSMHz SuperSP ARC II microproces­ "The ultimate in hot-vector technolot,'Y were sor from Mountain View, Calif.-based Sun the machines from [the extinct] Cray Computer Microsystems Inc. Cray recently demonstrated a Corp.," says Stephen Brobst, managing partner at SYSTEMS 4R-processor machine cranking away at a 1.6 ter­ market researcher Strategic Technologies and 1983-1989' abyte database. Cray claims the CS6400 has pene­ Systems, Cambridge, Mass. "They were phenome­ trated more customer sites in a shorter time than nally fast but they weren't economically feasible." any competing equipment. Cray's commercial cus­ Adds market researcher Burwen, "There won't be tomer base has grown so that it currently accounts any vector machines sold in another five years. for more than half of all Cray sales revenue. There's no future for those systems." "Even major parts of Cray believe [the tradi­ At first it seemed as though massively parallel tional supercomputer market] is dead," says con­ systems might in fact win out as mueh more cost­ sultant Brobst. "But they've done a terrific job engi­ effective than the vector machines. However, it neering the 6400 and moving into the commercial was much harder to create software for them. market. Only Cray could have done that." The software steadily improved but not as fast Thinking Machines has dropped out of the Iron as the microprocessors themselves. Suddenly, real­ business altogether. In October 1994 it reorganized world problems that people thought only Big Iron as a software-only company and promptly went on could solve-with either vector or massively paral­ to four consecutive profitable quarters. It apparent­ lel machines-were being tackled by machines ly has learned its lesson well. with fewer than 100 processors, and getting the job "The market spoke loud and clear, and told this done. The knotty problem of making software run company that building some of the world's fastest on a thousand microprocessors simultaneously is computers was not, by itself, enough to sustain often simply not worth tackling. "Microprocessors growth," TMC president and CEO Robert Doretti are getting so powerful that even a small number of admitted recently. "We took a hard look at our core them can handle problems that traditional super­ competencies and quickly realized that we had sub­ computers used to do," says Burwen. stantial expertise in the software that harnesses the In a recent study, the Palo Alto Management power of multiprocessor computers. We believe Group predicted that the market for parallel pro­ there's a huge untapped market for this capability, cessing systems would increase at better than a 40 and this is our strategic focus going forward." percent annual rate to about $14.3 billion in 1999. Approximately three-quarters of that amount will STAYING ALIVE come from commercial applications such as on-line There are, however, naysayers to all this negativity. transaction processing (OL TP), decision support "It's a big misperception that the high-end super­ systems (DSS) and multimedia. Science and engi­ computer market is dying," contendsCray neering will still be important markets but will Research President and COO Bob Ewald. Just last lose ground by century's end. November, Cray sold a top-of-the-line 32-pr~cessor The market will not be fed by Big Iron T90 supercomputer to Nippon Telegraph and machines. "There will be no l,OOO-processor Telephone Corp. (NTT).· According to Cray machines sold this quarter," predicts UC Berkeley's Chairman and CEO J. Phillip Samper, demand at Patterson. "Even machines costing a few million the end of the third quarter for the T90 line repre­ dollars are very unlikely to have more than 64 pro­ sented 45 percent of the company's backlog. Cray cessors. There's no demand for' the traditional big also claims to have $90 million in orders for its T3E supercomputers anymore." supercomputer, which isn't scheduled for delivery Companies are surviving by adopting the new until the first quarter of 1996. religion that says smaller can be better. Even And what does it say about the condition of a Seymour Cray's original company, Cray Research, market when Tera Computer, which since its once synonymous with Big Iron, has gotten the founding eight years ago has done absolutely noth­ drift. Its share of the traditional supercomputer ing but lose money-a mere $9.5 million-goes market is actually rising as competitors die off, public and oversubscribes the offering? "Big Iron is but the market is shrinking. A few years ago Cray not stone-cold dead," contends Patrick Grady, realized that it could no longer live on supercom­ senior vice president, corporate finance, at H.J. puters alone. In 1991 it created Cray Research Meyers & Co., the Rochester, N.Y.-based under­ Superservers after purchasing selected assets of writer of Tera's lPO. "It's contracting for the com­ Floating Point Systems. panies still able to play in it, but there's still plenty Last February, Superservers merged with two of business there. People are waiting for a new

44 UPS IDE JANUARY 1996 approach. Vector processing machines are long in the tooth. market dominated by nuclear weapons, weather forecasting Parallel processing machines are good but not exceptional. and a few industries, such as petroleum and aerospace, that The high-performance world needs a major breakthrough in needed major horsepower. programming." He believes that Tera is the company to do "Now we've come completely around," the adviser says. exactly that. "Big Iron supports a few niche markets, and the rest of the In fact, the technical supercomputing market could con­ world-because of the amazing increase in computer power tinue to stumble along for several more years. Bell paints a and distributed computing, now has on its desk machines picture with several types of surviving supercomputers, more powerful than Big Iron was 20 years ago. You use soft­ including Cray-style evolutionary supercomputers with ware tools and couple these together over corporate LANs multiple vector processors; multicomputers formed from and public data networks and it beats the shit out of Big Iron microprocessor-based workstations connected via in cost effectiveness." high-bandwidth, low-latency switches; and "mul­ But those niches may no longer be the breed­ tis," or multiple microprocessors connected to ing ground for entrepreneurs. "All the upstarts large caches that access a common memory via a that entered the supercomputer business are common . either dead or part of another company," market Bell also believes that a few trends could keep researcher George Smaby says. And those that the momentum for such machines going for a long aren't-MasPar, NCube (which is nominally inde­ time. For one, the govemment's "buy U.S." policy pendent but would disappear if Oracle Corp. is still alive, although not as visible as it was a few President Larry Ellison decided he had better years ago. For another, Cray, IBM and Silicon things to do with his fortunel, Meiko and Tera­ Graphics Inc. have large installed, loyal customer seem one phone call away from being eaten. bases for their supercomputers and super servers. John Toole, director of the National Coordin­ In fact, Intel Corp. announced the sale of a ating Office for High Perfonnance Computing and massively parallel machine just a few months ago. Communications, says that the huge amount of The Department of Energy's ASCI (Advanced capital required to play the high-performance game Scientific Computing Initiativel program is paying will prevent any startups from joining the roster. UNOW WE'VE COME COMPLETELY AROUND," A WHITE HOUSE ADVISER SAYS. HBIG IRON SUPPORTS A FEW NICHE MARKETS,"

$46 million for a 9,OOO-processor machine, which And Hillis adds that technology is changing so it el

JANUARY 1996 UPS ID E 45

it means things like darabase searches that technology and economics [Ook a and data mining for commercial users. rurn that lerr TMC with products that So what happened that brought weren't competitive," he says. TMC from the grand challenges of Specifically, the company was busy

the past [0 the throughput problems selling big, expensive systems at a time of the present? Six years ago, TMC, when the industry was discovering that then of Cambridge, MA, was one of it could do almost as well with inex­ the most glamorous o{ all the high­ pensive litde ones. tech firms along what was then called "And," admits Myczkowski, "there "AI Alley," a stretch of office buildings were also internal problems ... with not Elr from Massachusetts Institute management style." In particular, he Thinking Machines Returns of Technology. TMC had come into cites the company's long-standing Remember the hardware company being to produce the Connection belief that superior engineering alone Thinking Machines Corp.? The erst­ Machine, a supercomputer-like device was necessary to build a successful while maker of Connection Machines based on thousands of individual company, and rhat actually selling has tllrned irs feJClIs to sofrware and boxes was unnecessary and even vul­ inrrodllced GlobalWorks, which gar. "We had espoused the philosophy allows users ro take networks of that if we build it, they will come," exisring sysrems-parriclliarly Suns, he says. Rut, they didn't come. "We ClobalWorks runs on rap of should have become a marketing So/aris-and make rhem behave as driven company." though they were multiprocessor Rut management style problems devices. wt'nt even deeper for TMC. The com­ TMC describes ClobalWorks as a pany not only Eliled ro market its parallel-computing environmenr that product, it was very successful at alien­ rullS under UN IX. Whar ir runs on, ating potential customers. lr gained meanwhile, is the GlobalWorks server. the reputation for being among the This is a small high-bandwidrh ner­ most difficult companies in the indus­ work of Sun UltraSPARC-based sys­ try (() work with. tems--which themselves can be But, if pride goes before a hi!, then multiprocessor devices-housed in cab-­ sometimes {ails are instructive. inets that can hold up to 10 sysrems. Myczkowski says that the TMC of "The GloblWorks server is essenrial­ 1996 is a new company with a new ly a collection of Sun servers connected handle on things. The company, he by ATM or Fibre Channel," says Jacek says, has learned "that the marketplace Myczkowski, TMC's vice president of The G/oba/Works server from is governed by laws that are not the development and technology. Thinking Machines houses mu/tiple same as those of science." At the moment, GlobalWorks Sun systems. With the company's Thus, TMC's new direction. "We G/oba/Works environment, the device requires a GlobalWorks server ra run. can a/so function as a paralle/­ decided that our core competency But, says Myczkowski, there is no rea­ processing supercomputer. was (() leverage our software expertise son why you might not someday be and make it available on a wider set able to use the technology to link processors. Connection Machines of platforms," says Myczkowski. "We ATM networks of Suns into similar were thought to be the devices most decided to move it from the super­ clusters that don't reside in a TMC likely to give established high-perfor­ computer domain [0 a broader chassis. "In the-we hope-not too long mance computer vendors, such as marketplace. " rerm, you'll be able to choosf what Cray Research Inc., a run for their In other words, TMC's hardware you buy from us," he says. "It could money. was never exotic. Connection be hardware, or just software." Moreover, Connection Machines Machines were remarkable in appear­ GlobalWorks marks a significant were supposed [0 be the hardware that ance, but underneath they were pri­ shift for TMC, not only in terms of might, someday, achieve the Holy marily a collection of reasonably its product, but also in the way it Grail of computer science, rhe sen­ standard components. What was thinks about technology. "Parallelism tient machine. The company's motto unique about rhem was TMC's meth­ takes on a different meaning," says was "someday, we shall build a ods of making the components coop­ Myczkowski. "In the past, it meant machine which can think ... and it will erate. The company now intends to­ grand challenges. Today, it means a be proud of us." take those methods and sell them for throughput problem." Where before, As to what went wrong the first other devices, like workstations and it meant supercomputing and com­ time, Myczkowski cites various techni­ servers. "The network," says Myczk­ plex design or research problems, now cal and marketing reasons. "I think owski, "is a perfect environment for

12 SUNExPERT Magazine February 1996 , this sort of technology." likewise sought a market in the 19805. Myczkowski has an answer for that But can TMC corne back with a None of these came to much. Why, too. "I think we know how to do this. software-only solution? It's not, after then, does TMC think it can do what We have had the experience. We have all, alone in this. Products that allow others haven't? Myczkowski thinks had 12 years of dealing with parallel networked devices to cooperate have that the situation has changed. "In computing," he says. been around for quite a while. Over those days, there was a big difference In short, he says, "We've already a decade ago, Apollo Corp. (now a between the MIPS you had and the made all the mistakes." And that division of Hewlett-Packard Co.), for bandwidth of your network," he says. alone, he believes, is a significant example, had an Now, high-performance communica­ advantage, called Domain that allowed net­ tions technologies, like Fibre Channel, worked workstations to perform as make it possible for networks to a single device. behave as if they were parallel-process­ • During the days of lavish govern­ ing systems, says Myczkowski. ment funding on the Strategic But if that's true, doesn't TMC run Defense Initiative, there were several the risk of once again being outflanked federally sponsored efforts in the same by a commodity product? Won't every direction. And there were several vendor have a similar capability in its Transputer-based approaches that own product line?