Science & Technology: SUPERCOMPUTING
A SUPERCOMPUTER MAKER'S BRAVE OLD WORLD
Bucking the trend, Tera uses custom-crafted chips for extra speed
For computer scientist Burton J. Smith and his band of 65 employees at Tera Computer Co. in Seattle, the moment of reckoning is at hand. For nine years, they have slaved to build a new breed of supercomputer. Now, enough of its custom designed chips have been fitted together to determine whether Smith's theories will deliver on their promise. If so, Smith believes Tera could be the vanguard for future generations of supercomputers far faster than any yet built.
Trouble is, while Tera's engineers were laboring away, something happened in the outside world they hadn't foreseen: The supercomputer business changed radically. Systems built from ultrafast, proprietary brain chips, typified by Cray Research Inc.'s machines, account for a shrinking minority of the world's 500 fastest computers (chart). Instead, customers are gravitating toward supercomputers built from off-the-shelf microprocessors. While each chip is relatively slow, the computer gets its speed from harnessing dozens or even thousands of them.
DON QUIXOTE? This sea change in technology has been compounded by a flip in the market, with private industry surging ahead of the defense and intelligence communities as the primary customers. The government used to spend whatever it took to get more speed for designing nuclear weapons or breaking codes. But with the end of the cold war, the budget for "nukes and spooks" has shrunk dramatically, and several star suppliers have either winked out or moved on to other things, including Thinking Machines Corp. and Kendall Square Research Corp. (table, page 74).
Now, up pops Tera Computer, scorning the trend toward off-the-shelf chips with a new system based on a proprietary chip design. No wonder Smith is regarded by some as the Don Quixote of supercomputers. But Smith insists his quest will pay off "because we are addressing the right problems." These include a system design that, he asserts, will enable his computers to run at their peak speed roughly 95% of the time. That would be a huge advance, since supercomputers typically rev up to top speed just 10% or 15% of the time.
One key is "multithreading," a fancy term for a type of parallel processing. In most supercomputers with multiple brains, each chip handles no more than a handful of instruction streams, or threads, in parallel. The chips spend much of their time waiting for something to happen somewhere else in the system. One reason: The system can choke on all the control signals necessary to coordinate operations among its many processors. Programmers can reduce the communications logjam by tweaking the software, but that takes a lot of precious time.
In contrast, Tera's chips juggle 128 threads at once. So, if one thread runs into a roadblock and can't finish, the chip works on some other thread. Meanwhile, the processor parks the incomplete operation in a novel on-chip queue until the needed data arrives. "Our machine doesn't sit and wait around for any data," says Brian D. Koblenz, vice-president for software. And Tera has developed a so-called compiler that automatically reckons the best way to parcel out instructions.
Smith asserts that Tera's technology is crucial to overcoming the diminishing returns that now plague parallel processing. To add speed, suppliers shovel in more microprocessors. But that makes it still tougher to coordinate operations. As a result, new increments of speed get increasingly expensive. Tera claims its system does not suffer from diminishing returns. Thus, if one chip set can do 1 billion floating-point operations per second, or gigaflops, 256 chips would do 256 gigaflops. "Tera's approach is unique," says Jack J. Dongarra, a supercomputer expert at Oak Ridge National Laboratory. "That's what makes it so interesting from the science side."
To David Cutler, Microsoft Corp.'s manager of operating-system development, it's also interesting from the business side--so much so that he bought a stake in Tera and is on the board. Cutler says Tera's multithreading approach will be needed within five years for a new generation of desktops. "There is a real promise for the future," says Cutler, who headed development of the Windows NT multithreading operating system.
Some investors agree. After getting along for eight years on Pentagon contracts worth $18 million, Tera last year raised $9 million in an initial public offering and an additional $8 million in a private placement this past summer.
In part, that's an endorsement of Smith's reputation. In supercomputing's tight clique, Smith, age 55, is often mentioned in the same hushed tones as Seymour Cray, the father of supercomputing who died on Oct. 5. Robert R. Borchers, director of Advanced Scientific Computing at the National Science Foundation, says that if anyone can pull a surprise out of the hat, "it would be Burton Smith." Many Tera employees decided to risk working for a startup with an uncertain future just to sit at Smith's feet. "Burton is the reason why I came here," says Preston Briggs, a whiz kid who handled much of the compiler's development.
Smith's concepts first attracted attention in the 1970s, when he was chief scientist at Denelcor Inc. in Aurora, Colo. Denelcor spent 17 years and roughly $50 million trying to pioneer a supercomputer with a rudimentary version of multithreading. Although Denelcor sold a handful of systems in the early 1980s, it never posted a profitable year--and crashed in 1985.
Smith licked his wounds at the Supercomputing Research Center, a think tank in Washington, while continuing to nurse his dream. By 1987, he felt he had nailed down the missing details. So he hooked up with James E. Rottsolk, former vice-president for finance and development at Denelcor, to found Tera Computer, with Rottsolk as CEO.
Even some of Smith's admirers, though, worry that his brainchild has taken too long to arrive. If Tera had delivered on schedule two years ago, says Wayne Pfieffer, associate director of the San Diego Supercomputer Center, "it would clearly have been an exceptional system." Now, he says, "it's not clear how it will stack up to the competition."
Tera not only designed a custom microprocessor--at a time when standard models have moved to the fore--but it's using gallium arsenide to make it. That substance is inherently at least five times faster than silicon, but low production volumes will make the chips more costly. "You cannot fight the market and win," says Larry L. Smarr, director of the University of Illinois' National Center for Supercomputing Applications. Smarr is buying only computers made with standard chips--mainly from Silicon Graphics Inc. and Hewlett-Packard Co.
Indeed, standard chips are sweeping the field because of their huge advantage in price for performance. "Fifteen years ago, a Cray-1 cost $18 million and gave a little less performance than a Pentium Pro does today" for about 0.01% of the cost, says Edward A. Masi, an Intel Corp. vice-president and a former Cray manager. Even the world's biggest and costliest computer, a $110 million behemoth that SGI and Cray will deliver to Los Alamos National Laboratory in late 1998, will be powered by microprocessors--3,072 to be exact. These will be the same MIPS Technologies Inc. chips that go into SGI's workstations. Together, they will spit out better than 3 teraflops, or 3 trillion calculations a second.
THE NEXT PLATEAU. Tera had hoped to live up to its name and at least bid for the job of building a teraflops system for the Energy Dept. But without even a prototype to back up its claims, it had to walk away. Soon, though, Tera expects to begin offering a small, 16-processor system for $12 million, then add larger systems of up to 256 processors.
The ultimate payoff, Smith believes, will come when supercomputers push toward the next plateau in speed--petaflops, or quadrillions of calculations per second. That will require a fundamentally new approach, and Smith is certain that Tera's system points the way. "Tera is either the last of the old breed or the first of a new breed," says supercomputer analyst Gary P. Smaby, president of Smaby Group Inc. in Minneapolis. One thing is clear: Burton Smith and Tera merit watching.By Seanna Browder in SeattleReturn to top