Still, some people in the industry saw a bit of irony in the Government's ponying up the $45 million to develop Intel's new machine...
By LAWRENCE M. FISHER
Published: September 11, 1995
New supercomputers have always been a bit like new jumbo jets: without an initial customer and the big opening order it brings, they never take off. But while Boeing can turn to any number of domestic or foreign airlines for that crucial first check, there has always been just one qualified buyer for computing's ne plus ultra: the Federal Government.
So the announcement last week that the Energy Department would finance the creation of the Intel Corporation's next stab at the world's fastest computer was very much in keeping with tradition. The big difference was that "Pay to the order of" was not followed by "Cray Research Inc.," the leading producer of supercomputers for the last two decades, but by Intel, best known as the maker of the microprocessors at the heart of most personal computers.
That the role of supercomputer leader should be claimed by a maker of personal computer chips is a technological upheaval worthy of the overused phrase "paradigm shift," but it should not be surprising. Hit simultaneously by the end of the cold war and a global recession, the traditional supercomputer market has been in decline for several years. Both Seymour Cray and Steve Chen, the competing geniuses of Cray Research, saw their subsequent ventures fail because no one stepped up to be the first to buy.
At the same time, the personal computer industry has boomed, and technology leadership has simply followed the dollars. While Cray's machines depended on relatively few but vastly powerful processors, Intel's supercomputer will use 9,000 off-the-shelf microprocessors arrayed together in a configuration designed to achieve what is known as massively parallel processing.
"It used to be that big machines pushed the performance envelope," said Ed Masi, general manager of Intel's supercomputer business in Beaverton, Ore. "Today the desktop pushes all technology. That's where all the revenue growth is, where all the profit is, so that's where all the research and development is."
Still, some people in the industry saw a bit of irony in the Government's ponying up the $45 million to develop Intel's new machine, which will be based at the Energy Department's Sandia National Laboratories in Albuquerque, N.M. A previous Intel machine, also at Sandia, already claims the world speed record, but that computer and all of the company's current offerings are based on a microprocessor, the I860, which is not used in personal computers and is considered, even by Intel, to be an obsolete chip. The new machine will be based on the P-6, the successor to the Pentium, which is due out later this year.
"The I860's performance was out of line with the performance of other vendors," said Omri Serlin, president of Itom International, a market research firm based in Los Altos, Calif., that follows the high performance computer industry. "They were in a quandary for what to do. Now they've got a sugar daddy to finance a machine based on the P-6."
With the current Congress's focus on reducing the budget deficit, an Energy Department subsidy for one of the world's most profitable companies might raise a few eyebrows.
But supercomputer experts said that the estimated $45 million cost should not be considered the purchase price of a computer that can perform more than a trillion floating point operations a second but the expense of solving a particular problem -- in this case, simulating a nuclear explosion.
(This capability, known, uneuphoniously, as a teraflop, involves the ability to perform simultaneously hundreds of billions of computationally intensive operations, and is used as a standard reference for comparing supercomputer performance.)
Mr. Masi expressed hope that the French Government might be Intel's next client.
"Here is not an announcement of a new machine but of an end-user, the Government, with a particular problem of interest to society, saying we can justify the expenditure of this money to solve this problem in the time frame we need to," said Sidney Karin, director of the San Diego Supercomputer Center.
Time is the critical component, Mr. Karin said, because the history of computing shows that performance always increases and cost always falls if one waits long enough. "I think that if you did nothing, sooner or later there'd be a teraflop on your desktop," he said. "Whatever is happening at the high end will and should arrive on your desktop 5 or 10 years from now."
In the past, the high-end machines that Cray sold first to the Defense Department, the National Security Agency or some other Government body soon found buyers in the commercial market. The Ford Motor Company, for example, is one of the biggest Cray users and the seamless body of the new Taurus would not have been feasible without lots of supercomputer time. Aerospace, oil discovery and even pharmaceuticals manufacturers have been strong supercomputer buyers.
It is less clear that massively parallel machines like Intel's 9,000-chip behemoth will find nongovernment markets. While each succeeding generation of Cray computers maintained software compatibility with its ancestors, massively parallel machines require an entirely new approach to writing code that has slowed their adoption. The high-end supercomputer market may be shrinking, but Cray retains a robust 70 percent share.
"There doesn't seem to be that much interest in the commercial marketplace for the high end of parallel processing," Mr. Serlin said.
On the other hand, massively parallel machines are inherently scalable, meaning that the same design could be used with 4,000 or 400 or 4 microprocessors, providing incrementally lower performance but at lower cost. And it is the so-called midrange of supercomputers, machines priced from $2 million to $5 million, and the so-called low end,at $1 million and below, that are experiencing rapid growth.
The dominant player in the midrange and below, and to some extent the creator of the small supercomputer, has been Silicon Graphics Inc. But Cray introduced its own midrange machine in March, which has turned out to be its fastest-selling product ever. "We launched that machine with 120 advance orders," said Steve Conway, a Cray spokesman. "It took us 10 years to sell 120 conventional supercomputers."
Still, this is the market Intel thinks it willcommand. "We will sell pieces of the larger work, much smaller systems, even down to work stations," Mr. Masi said. "I believe the P-6 will take over that marketplace."
