Issue No.04 - April (2007 vol.40)
Published by the IEEE Computer Society
David Alan Grier , George Washington University
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/MC.2007.146
At one time, supercomputing, a term that was synonymous with the Cray-1, was considered to be the best deal in town. However, by the 1990s, manufacturers began to question the need to continue the development of high-speed computers, turning their attention instead to the potential for developing a "high-capacity and high-speed national research and education computer network."
Robb walked out of his office with a grin on his face and a letter in his hand. "If business gets any better," he announced, "I'm going to quit this dump and join the Foreign Legion." His staff, unsure how to interpret this announcement, stayed at their desks and waited for an explanation. They knew that business had not been good, and they knew too well that Robb had an unusual way of expressing himself. Only the week before, they had heard him suggest that they might try to market their computer next to a subway station. "We'll put up a big sign," he had said. "One minute of supercomputer time: $1.00. It will be the best deal in town."
No one had laughed then and no one was prepared to laugh now. The group had faced disappointment after disappointment as they worked to develop and market a high-speed scientific computer, one of a class of machines that had been termed "supercomputers." Delay had followed delay. A simple design had grown complicated. Marketing opportunities had passed while the group stood silently on the sidewalk.
Even the first demonstration had been a disappointment. On that morning, the newly born supercomputer had failed to respond for nearly two hours. The control lights flashed, the disk whirred, but nothing happened. When the machine finally accepted a program, it had limped through the calculations. Robb had retrieved the program output when it appeared at the printer, scanned it for a moment, and then solemnly addressed the assembled group of engineers and managers.
"Colleagues," he had said, "we have spent $25 million to breed a champion race horse." He looked down at the numbers again and then stared at the crowd. "But we are no closer to that goal than if we had begun by cloning a giant whale."
With such a history, Robb had little credibility as the bearer of good news. After the machine was operational, he had made at least three dozen marketing presentations and had yet to sell a single machine. Sensing the distrust among the assembled group, he paused. "What do I need to do?" he asked. "Should I run buck naked through the halls yelling 'Eureka'?"
The thought of the burly physicist running unclothed through the building brought a smile from one corner of the room, a giggle from another, and a gentle sigh from Emily at the secretary's desk.
"We have victory in the Pacific," Robb announced. "We will sign the instruments of surrender tomorrow at noon. We have sold a supercomputer to Japan!"
The room filled with the cry of voices, the slapping of hands, and the suggestion that business cease for the day so the staff could retire to the "Old Automaton," a cramped and decaying bar across the street.
The sale was a major accomplishment as the market for scientific computers was crowded with six companies. Each claimed to have the best machine, the quickest chips, the cleverest software, the fastest design. Each had invested millions of dollars in its supercomputer and each believed in the value of its investment.
Computers that were Super
Once upon a time, every computer was a supercomputer. The ENIAC bore this title, even though it was the fastest computer by default. The Univac I, which was far from the fastest machine of its age, was once characterized as a "super-duper electronic brain, which can think in terms of a couple of quintillion mathematical problems at one time." For more than two decades, eager sales people applied the term to describe any innovative machine. However, by the 1970s, the word supercomputer had narrowed in scope so that it referred to scientific computers. The machines most commonly associated with supercomputing were those designed by Seymour Cray.
Cray began designing computers in 1950, when he was a young engineer with Engineering Research Associates. Although his first designs, which were created for the US Navy, were not especially fast nor did they contain the features that marked his later machines, they were solid, lean creations.
Cray began focusing on high-speed scientific machines when he designed two computers for Control Data, the 6600 and the 7600. With these machines, he learned how to make the most of every element in the system. Both were highly successful and were adopted by some of the most demanding research laboratories, including the Navy's weather research facility and the nuclear weapons laboratory at Los Alamos.
A special machine
In 1976, the term supercomputer was synonymous with the Cray-1 computer, which was released that year. There were another half-dozen operating computers that might have had some claim to being super, but the Cray-1 demanded that role for itself. It was extraordinarily fast for its time. When all processors were fully operating, it would perform 80 million floating-point operations per second.
The Cray-1 also looked like a special machine. Shunning the traditional boxes of ordinary computers, Cray built a machine that looked like a cross between a prop from the first Star Wars movie and a couch from the lobby of a Parisian hotel. The central processor was a wide cylinder that stood six and one-half feet tall. Around the base of the cylinder was a curved, black vinyl bench that hid the machine's power supply.
The first organizations to purchase a Cray-1 were the facilities that had used Cray's earlier machines: the Los Alamos weapons laboratory, the National Center for Atmospheric Research, and the European Centre for Medium-Range Weather Forecasts. These laboratories were more than willing to bear the full cost of a new computer, which included not only the price of the hardware but the investment in software. The first machines came with a simple operating system, an assembler, and little else. The first users had to debug the software, write a Fortran compiler, and convert their programs.
The development of software required the programming staffs to invest considerable amounts of time in understanding the Cray-1's design. As they worked with the machine, they came to realize that few programs would be able to command the Cray's full power without substantial modification. This issue came to be known as "the dusty deck problem." How did you take an old program that had been stored on punched cards and make it run fast on the machine?
To solve the dusty deck problem, a programmer had to understand the vector processor. This processor performed computations on large data sets and worked in parallel with the rest of the machine. It was both the machine's blessing and its bane. When fully employed, this processor drove the Cray-1 to the limits of its performance. If it was idle, the machine limped along, barely faster than some of the data processors of its age. When running in this more inefficient state, the Cray was still a fast machine, but not fast enough to justify its multimillion-dollar price tag.
At Los Alamos, the programmers slowly learned how to adapt programs to keep the vector processor fully employed. But the adaptations were not always the most natural versions of the program. They often had to invert a program—rewrite it so that the machine would do the calculations in a new order. "Think vector" was their oft-repeated motto. But they found that vectors were not always the easiest things to keep in thought. All the early users of the Cray found that they had to rethink and reorganize their programs.
"Originally, Seymour Cray had set out just to build 'the fastest computer in the world' and let the users worry about software," observed the historians who chronicaled the development of the Cray-1. But the early customers resisted this idea. In particular, the National Center for Atmospheric Research "decided not to buy the machine unless [Cray] supplied system software as well."
Cray did not like this idea. "I would rather use the corporate resources to explore and develop newer and more unique computing equipment," he wrote in an annual report. He acknowledged that this strategy would tend to limit growth because it would mean that his company would move "out of market areas when competitive equipment begins to impart course sales." However, Cray argued that it was in the best interests of his stockholders "to limit growth in this manner and maintain a good profit margin on a smaller sales base." Nonetheless, he eventually accepted the idea that his firm would have to develop software for its creation.
Challenges for the prize
At first, Cray built four computers a year. He had no real competition and plenty of potential customers. He peddled his machines to the government laboratories, large engineering firms, and weather agencies. Slowly, his success began to attract some competitors: Texas Instruments, Control Data Corporation, Burroughs. However, none was able to capture much of the market as Cray's company had become the dominant supplier. "I always found that Cray salesmen were very friendly," Robb would later recall. "They didn't believe that any firm like ours could compete with them."
In fact, the three American firms proved to be poor competitors. Control Data was the most successful, but it was only able to sell a handful of machines. Texas Instruments sold one. Burroughs delivered none. Of the 61 supercomputers that were operational in 1982, Cray had built the majority.
The supercomputer market was starting to expand rapidly as laboratories outside the US began purchasing high-speed computers. Seven of the 61 supercomputers of 1982 were located in England, six were in Germany, four in France, and two in Japan.
While Cray applauded these sales, other computer scientists viewed them with alarm. There "is the worry that American basic research and engineering science are going to fall behind," argued one scientist, "especially in light of the progressive national policies that Germany, France, Great Britain, and Japan have concerning the exploitation of the supercomputer's potential—including policies to make supercomputers available to their university researchers."
Over the next six years, the number of installed supercomputers increased nearly sevenfold to 409 machines. Only half of these machines came from Cray. Some came from new American firms, but a growing number came from three Japanese companies: Fujitsu, Hitachi, and NEC. All three companies had taken market share from Cray. An NEC machine had surpassed the Cray computers, acquiring the title of the world's fastest computer.
The rise of the Japanese supercomputer industry troubled the leaders of the American computer business. When a "group of leading computer science experts in the United States gathered to assess Japan's progress," reported one journalist, "they emerged envious and more than a little scared."
This was one more piece of bad news in a time of great anxiety for American firms, which had seen large parts of their computer hardware business move to Asia. Japanese firms had become the dominant makers of memory chips. Fujitsu was trying to purchase Fairchild Semiconductors, the company that supplied chips to Cray.
It isn't always easy to identify a trade war. You don't see troops massing on the frontier, weapons being stockpiled, or soldiers being trained. American and Japanese firms confronted each other in a number of markets, but few conflicts were more visible than the field of supercomputers. "It is a matter of national pride and national identity," wrote one commentator. "The United States took 20 years to build its supercomputer industry. Japan has built its industry in six."
A Changing Time
In the late 1980s, the American government began to study Japan's position in the supercomputer market and consider how the three leading firms were able to sustain their place in the in the industry.
To many observers, including most of the management of Cray's company, the Japanese government seemed to be unfairly subsidizing supercomputer sales. They claimed that Japanese firms were dumping machines, selling them to American institutions at a price that was substantially below the cost of development. The American government accepted these charges and demanded that Japan restrict the export of its machines. Japan took a different point of view and rejected this demand.
The two countries feuded over supercomputers for five years and were unable to resolve the disagreement. After exchanging charges and countercharges, they signed an agreement on supercomputer trade, but few people thought the agreement did any good. The head of Cray complained that the agreement only allowed the Japanese firms to "see all the contracts that we are excluded from."
As the two countries feuded, many of the companies in the supercomputer business struggled. Many questioned whether they should invest more of their time and treasure in the business.
After a decade in supercomputer sales, Robb decided that he needed to seek a better opportunity in some other part of the industry. He felt he had never earned enough from his supercomputer royalties. He had to support a wife, a daughter, a horse, and a country home. He needed to sell a product that had a larger market.
Robb's departure was a subdued affair, a lunch that was more awkward than joyful. His colleagues were not sure if they should be wishing him success in his new job, which was with a Japanese firm, or if they should ask a blessing of him. He was not in a philosophical mood. "One day I had to ask if I had been conned," he said. "I needed to know if I should devote the rest of my life to supercomputers. I could not answer that question, and so I decided that I had to find a job that would better support my family."
The End of an Era
Robb was not the only individual to question his support of the supercomputer industry. In the first years of the 1990s, both the US and Japan questioned their need to continue the development of high-speed computers.
In the US, the questions were driven by the end of the Cold War, the four-decade-long conflict with the Soviet Union. No longer facing an aggressive enemy, the US reduced its support of military research—research that had sustained several supercomputer firms.
In Japan, the doubts were fueled by a recession that undermined the economy. In a six-month period, the Japanese stock market lost 38 percent of its value. The country could no longer afford to support the development of computers that did not generate revenue.
The end of the supercomputer era was marked by the decline of Cray's company. Its name was no longer synonymous with the supercomputer. In the spring of 1993, the company attempted to sell its top machines to seven different laboratories. Only one agreed to purchase a Cray.
During his career selling supercomputers, Robb had used many metaphors to describe high-speed machinery. As he left the company, he returned to a favorite. "They are a like a Ferrari sports car," he said. "They're fast. They're beautiful. They cause men to go temporarily insane. They cost a fortune and spend 90 percent of their life in the shop being adjusted."
"No one really wants to own a Ferrari," he concluded. "They want to have a wealthy uncle buy one and then lend it to them on weekends. That would be the best deal in town."
If there was one event that ended the age of Seymour Cray, it was the High-Performance Computing and Communication Act of 1991, the legislation championed by then-senator Albert Gore Jr. The bill praised the value of supercomputers. "Advances in computer science and technology are vital to the Nation's prosperity," Gore wrote.
Understanding that such machines were as seductive as an expensive sports car, Gore argued that most of the country's researchers did not need such a machine. Instead, the US government should have a "high-capacity and high-speed national research and education computer network" to "provide researchers and educators with access to computer and information resources."
Gore wanted to put a sign out on the sidewalk that read, "One minute of supercomputer time: $1.00." It would be the best deal in town.
David Alan Grier is the editor in chief, IEEE Annals of the History of Computing, and the author of When Computers Were Human (Princeton University Press, 2005). Grier is an associate professor in the Center for International Science and Technology Policy at the George Washington University. Contact him at firstname.lastname@example.org.