February 9, 1998
A Maverick Builds a New Supercomputer in a PC World
By JOHN MARKOFF
EATTLE -- In the arcane world inhabited by the builders of the world's fastest computers, Burton Smith is the last man standing.
A large rumpled man who speaks at a measured pace, Smith is a throwback to an era when men like Seymour Cray, Steve Chen and H. Danny Hillis competed to design the fastest, most beautifully handcrafted supercomputers and assembled them painstakingly, one by one.
But Cray's first company, Cray Research, was acquired by Silicon Graphics in 1996, and his second company, Cray Computer, foundered when he died of injuries suffered in an automobile accident in 1996. Hillis' Thinking Machines Corp. and Chen's Supercomputer Systems Inc. both failed when corporate sales developed slowly.
Credit: Robert Sorbo Burton Smith with a partially assembled computer.
Today, most of the world's high-performance computers are assembled Lego-style as hundreds or thousands of off-the-shelf microprocessor chips from a handful of companies including Silicon Graphics, Intel, Sun Microsystems, or IBM are hooked together.
The high-performance computing market today is increasingly dominated by these massively parallel, or MPP, computers, which harness hordes of so-called killer micros that achieve supercomputing speeds by breaking large problems up into tiny chunks that can be solved simultaneously.
Surprisingly then, in his offices here overlooking Lake Union, Smith is on the verge of achieving something that the skeptics thought impossible: creating a new supercomputer company based on a new approach to parallel processing.
After a decade in the laboratory, Smith achieved a measure of vindication recently when the San Diego Supercomputer Center installed the first system to be shipped by his company, Tera Computer Co. The $6.1 million system was underwritten by the Defense Department and the National Science Foundation.
Like many start-ups, Tera is still thought of as a huge gamble even by those who are rooting for its success. Nothing has come easily for the Tera engineers. More than a year late, only one processor of what may soon be a 16-processor supercomputer is working today. And Tera's engineers are still wrestling with last-minute bugs.
Still, no one is ready to count Burton Smith out, and there are many who believe that his machine will still shake up the computing world when it is complete.
"I admire his persistence," said Hillis, who is now a research fellow at Walt Disney Co. "Burton is the last independent supercomputer designer."
Supercomputers, defined as the fastest computers available at any given time, were first used for designing nuclear weapons and predicting weather but have increasingly been used in applications ranging from designing drugs to simulating car crashes. However, with the end of the Cold War and the decline of government financing for supercomputing and the simultaneous surge in the power of desktop computing, innovations increasingly come first from the PC and consumer electronics worlds.
That has led to a decline in research into new computer designs. The number of companies trying to build the most powerful 64-bit processors needed to drive the next generation of computing has dwindled to a handful.
With the death of Cray in 1996, many computer designers felt that the final chapter had been written for an industry that only a decade before was widely seen as crucial to the nation's technological survival.
But the people making those predictions did not reckon with Smith. "I'm feeling very good about where we are now in spite of the fact that most people in the computer business think we're out of our minds" he said.
Still, despite his optimism, there are people in the computer industry who are deeply skeptical about Tera's chances for survival in a computer world that is increasingly homogeneous and dominated by a handful of multinational giants.
Smith is one of the few remaining independent innovators in computer design.
"The supercomputer industry has been transformed in the last five years," said Larry Smarr, director of the National Center for Supercomputer Applications in Urbana-Champaign, Ill. "Only market-driven companies with a strong desktop business are going to be able to survive."
C. Gordon Bell, a well-known computer designer, is fond of calling the new Tera system "Burton's folly."
Most executives in the computer industry think that almost all of the fastest computing systems in the first decade of the next century will be based on the 64-bit Merced microprocessor chip being designed by Intel and Hewlett-Packard.
But there is a significant market for high-performance computing both in the technical marketplace and for scientific and corporate research. Industries from airplane manufacturing to banking could use ultrafast commercial computers for applications like reservation systems and credit card processing.
International Data Corp. estimates that the market for technical high-performance systems reached $3.57 billion in 1997, while the Gartner Group is predicting that the total market for high-performance computers will reach $5 billion this year. Most of those systems will be sold by Silicon Graphics, IBM and Sun Microsystems.
Compared with the sales of these giants, Tera's will be tiny this year. Salomon Brothers estimates that the company will have $49 million in revenue in 1998, rising to $106 million in 1999. It predicts that the company will be profitable this year.
It is Smith's ability to balance hardware and software that has most impressed many of the leaders in the supercomputing world. Last year, a single processor version of the Tera machine set a world speed record for sorting integer numbers.
The design is based in large part on Smith's rejection of the idea that any computing problem can be solved best by simply assembling ever-larger numbers of microprocessor chips. In fact, he thinks he has found a way to overcome the Achilles' heel of today's parallel supercomputers, what computer designers refer to as memory latency -- a measure of the time wasted in a supercomputer while microprocessors wait for new data.
That stumbling block has created a huge gap between what researchers call "theoretical peak performance" and "sustained performance" in measuring computer speed. Connecting hundreds of ultrafast microprocessors does not guarantee supercomputing speeds. Smith's system is intended to work in a fashion similar to a secretarial pool. As soon as a processor finishes a task it picks up a new one from a large pool of waiting tasks, ensuring that all of the processors are efficiently used.
The Tera approach, if it proves successful, will affect the entire field of computing because an increasing problem at all levels of computing is the growing imbalance between processor speed and memory speed. Even desktop PC microprocessors spend much of their time waiting for new data. The Tera design largely solves this problem.
A second hurdle that has plagued the parallel computing world also works in favor of the Tera supercomputer. Earlier generations of parallel computers have posed tremendous programming challenges for software designers accustomed to designing their applications for single processors.
Converting programs, which may consist of millions of lines of code, to run on machines with large numbers of processors has proved to be a greater challenge than was originally thought.
In contrast, Smith thinks that in the $5 million to $40 million Tera machine -- depending on the number of processors ranging from eight and up -- he has discovered a general-purpose supercomputer that can automatically break thorny scientific problems into smaller pieces that can be solved simultaneously.
That approach may prove to be a huge advantage for the Tera compared with other supercomputers that may have higher peak speeds. Smith thinks that it will win commercial orders this year from companies like airplane and automobile manufacturers that run huge simulation programs for each new model. He hopes for other sales to government intelligence agencies and research laboratories.
Indeed, Tera was originally supported by the Defense Advanced Research Projects Agency, the Pentagon's research arm, when Smith moved from the Supercomputer Research Center in Maryland, now the Center for Computing Sciences, to Seattle to found Tera in 1988.
Smith, who fell in love with computer math as a graduate student at the Massachusetts Institute of Technology during the 1960s, spent three years at the center, a contact point between the academic world and the secretive National Security Agency.
He went there after the failure of an earlier supercomputer company in Colorado, Denelcor. Smith was the architect in the early 1980s for the Denelcor HEP, the first commercial parallel computer. Denelcor went bankrupt in 1985 when its machines failed to find a significant commercial market.
Undaunted, he has raised more than $47 million in the development of the Tera MTA supercomputer, and he is planning to add a second processor to the first machine at the San Diego Supercomputer Center later this month.
Among the small fraternity of scientists and engineers who make up the nation's supercomputer cognoscenti, he has many backers.
"Burton has a deep understanding of what limits the performance of today's high-performance computers," said Robert R. Borchers, director of the division of advanced scientific computing at the National Science Foundation.
It was the belief that Smith is one of the few remaining independent innovators in computer design that led a senior official in the U.S. intelligence community to approach Sid Karin, the director of the San Diego Supercomputer Center, to ask him to help support Tera.
"He said, 'We don't have a lot of innovative architects like Burton Smith and Seymour Cray, and they need to be nurtured and supported,"' Karin recalled.
Following are links to the external Web sites mentioned in this article. These sites are not part of The New York Times on the Web, and The Times has no control over their content or availability. When you have finished visiting any of these sites, you will be able to return to this page by clicking on your Web browser's "Back" button or icon until this page reappears.
- Silicon Graphics Inc.
- Sun Microsystems, Inc.
- IBM Corp.
- Intel Corp.
- San Diego Supercomputer Center
- Tera Computer Co.