Super-computer race hit by costly development

Press, 22 September 1987, Page 26

Super-computer race hit by costly development

By.

LAWRENCE EDELMAN

•f Reuter (through NZPA) New York The international race to develop a new generation of super-computers will go on in spite of a leading United States manufacturer’s abrupt decision to abandon its most advanced project on grounds of cost, industry analysts say. Cray Research Inc., said it had dropped one of three programmes launched to make a super-computer 100 times more powerful than any yet built It said the programme would have been too costly to complete and that the firm would focus on other, less ambitious plans. The cancelled computer was being designed to solve complex problems at lightning speed by using up to 64 processors to split the work into dozens of smaller pieces. That concept known as “highly parallel processing” is a big leap from the four processors used in

the most powerful systems Cray now sells. "Cray’s decision was not necessarily a statement that parallel processing is not the way to go,” said Jeffrey Canin, an analyst with Hambrecht and Quist in San Francisco. “It was a statement that it doesn’t fit with their strategy,” he said. Cray and other supercomputermakers have turned to parallel processing to meet the needs of super-computer users who are clamouring for faster and more versatile machines. These users, mainly Governments, universify researchers, and corporations, use supercomputers for jobs such as designing aircraft and spacecraft and forecasting global weather trends. The decision to end the programme was clearly a setback for Cray’s development efforts, analysts said. It comes as the company, which has dominated the field, faces growing competition from United States and Japan-

ese rivals. It did not mean that radically faster super-computers were unobtainable, they said. Indeed, Steve Chen, chief designer of the computer, intends to seek financing to continue his work elsewhere. Meanwhile super-computer-makers said Cray’s action in no way changed their plans to build parallel processing machines. “We absolutely believe that better performance in the future will be gained by using more and faster processors,” said Norman Dawson, of Chopp Computer Corporation, which is finishing development of an eightprocessor super-computer. “We believe that Mr Chen was doing good work,” Mr Dawson said. Anaylsts said Cray ended the programme because costs had got out of control, not because the goal was impossible. Cray said it would have taken more than SUSIOO million ($162 million) to complete, twice its original estimates. Cray instead will focus on development of two

super-computers that use fewer processors but can still crunch numbers at incredible speeds. Cray’s Y-MP model, also designed by Mr Chen, will have eight processors. It will be ready sometime next year. The Cray-3, created by Cray’s founder, Seymour Cray, will be equipped with up to 16 processors and is scheduled to appear in 1989. It will be 100 times faster than the Cray-1, the first Cray super-computer introduced in 1976. "In going for 64 processors, Mr Chen is biting off the whole enchilada,” said Peter Patton, chairman of the Consortium for Supercomputer Research in Minneapolis. Mr Patton, comparing Mr Chen, aged 43, to John Rollwagen, Cray’s chairman, said: “He has a very good vision of the future. He is telling us what computing will be like in 1995. Rollwagen is running a computer company trying to make it to the 19905. “It’s really a question of far-term versus near-term goals,” Mr Patton said.

Regardless of the number of processors, supercomputer designers must overcome several obstacles if they are to achieve big performance advances. Although they have developed new processors that can handle data at high speeds, designers are still struggling to find efficient ways lo connect them. Software rn’ist also be written that divides a problem into pieces solvable by the processors. Cray’s X-MF machine, which has four processors, still solves many problems with a single processor, although four separate problems can be worked on at one time. - Finding answers to these and other problems will take time. “You just can’t push developmental fronts too fast,” said Steve Hemping of ETA Systems, a unit of Control Data Corp, which is posing the first real challenge to Cray. “They have to come in their own time,” he said.