Alternative to Costly Supercomputers : Technology: AMT of Irvine believes that the “massively parallel processing” field has gained legitimacy but needs compatible software to get into the marketplace.
Active Memory Technology Inc. has made its mark with super-fast computers that can perform dizzying tasks involving millions of calculations such as creating a precise, three-dimensional image of a DNA molecule, or searching for booby traps in a mine field.
But Geoff Manning, chairman and chief executive of the small Orange County company, doesn’t want AMT’s computers to be locked in guarded research laboratories and worshiped like a national treasure. Instead, Manning aims to sell to mainstream corporate users at a price that delivers more bang for the buck than some supercomputers.
Manning’s challenge is to wait until corporate America and other big technology users such as hospitals, universities and the military are ready to develop the software that could take full advantage of his machines.
After 3 1/2 years and $17 million in financing, AMT is a leader in the so-called “massively parallel processing” field with 1989 sales of $7 million, up from $6.5 million in 1988. But the company isn’t profitable yet and has been troubled by management turnover: Its former chief executive quit four months ago, forcing Manning to step in to fill that job.
And several other parallel computer companies with similar goals as AMT have given up waiting for the market to develop and have gone out of business.
“Aficionados would say massively parallel processing has taken a long time to develop a market,” Manning said. “But I believe it is now accepted as the only way to move computing forward.”
AMT, Thinking Machines of Boston, Intel Scientific Computers of Portland, Ore., MasPar in Sunnyvale and a half-dozen other makers of massively parallel computers are taking a chance that computing’s future lies in their particular approach to making faster machines.
The traditional “serial” computer, which has been the primary computer architecture for three decades, has a smart central processor that receives data in a single stream from main memory and solves one problem one at a time.
But as performance demands heighten, serial computers are slowed by something called the “Von Neumann bottleneck,” named after mathematician John Von Neumann, who dreamed up the idea of a central processor in 1946. The bottleneck, which results as data flows to a computer’s central processor, is like a traffic jam on a bridge with a single toll booth.
Massively parallel processors, on the other hand, use anywhere from 1,000 to 64,000 simple processors that together work on different pieces of a problem. The process more closely mimics the way the human brain works, researchers say, and would be like adding thousands of toll gates to that toll bridge to unlock the traffic jam.
Breakthroughs in both component manufacturing and computer design have enabled AMT to build a machine that can outperform minicomputers at the same price. While parallel systems cannot beat conventional computers in some tasks, such as balancing a checkbook, they perform well in tasks with intensive amounts of uniform data to process, such as analyzing a satellite image.
While parallel processing computers have been around since the 1970s, they have been largely ignored for commercial applications because of the difficulty of developing software to run on the machines, said Chris Willard, a supercomputer analyst at Dataquest, a San Jose market research firm.
But as parallel processors become faster, and traditional supercomputer designers run into physical barriers that limit their performance gains, the parallel approach is gaining legitimacy, Manning said.
That view was underscored by reports out of Tokyo last week that Japan’s Ministry of International Trade and Industry is planning a major research program to develop massively parallel processing. With such an effort, Japan’s leading supercomputer makers--Hitachi Ltd., Fujitsu Ltd. and NEC Corp.--could start playing catch-up with the American parallel computer makers.
Japan’s moves are “a sign of legitimacy, or rather a reaction to the legitimacy of applications that have been developed,” said Danny Hillis, founding scientist at Thinking Machines. While some believe that the Japanese face a daunting task in catching up on a decade of software development in the United States, Hillis takes no comfort in that argument and says U.S. companies will face intensified pressure to find commercial markets.
Manning agreed, noting that “once (the Japanese) start, they go very fast.”
The United States, Hillis said, staked out a lead in parallel computing with help from the Pentagon’s Defense Advanced Research Projects Agency, which began funding research into the field in 1984.
In contrast to Thinking Machines, AMT isn’t aiming to build the world’s fastest computer for the scientific community. Instead, AMT is focusing on users in the defense, commercial and academic fields.
Since 1987, AMT has introduced three generations of its parallel computers. In late April, AMT unveiled the DAP/CP8 series, which is 40 times faster than the company’s first machine. To keep pace with competitors, Manning wants to improve speeds by a factor of 25 by 1997.
AMT’s machines are connected to Sun Microsystems Inc. or Digital Equipment Corp. workstations and employ between 1,024 and 4,096 processors. They sell for $120,000 to $415,000 and reach speeds in the low-end supercomputer range.
The medical school at Washington University in St. Louis has purchased two AMT computers to help it speed up the process of scanning images to detect tumors in humans. Instead of waiting days for a minicomputer to generate sharp scanner images, the AMT machines complete the task in minutes, said Robert Morley, an associate professor of electrical engineering.
“A supercomputer would have cost millions of dollars, and that would have been out of the question,” Morley said.
Willard, the Dataquest analyst, said parallel processing has potential in commercial markets. He said the market for such computers has sprouted from $8.5 million in 1984 to $75 million in 1989.
Despite the progress, “this is an industry waiting for the users to catch up with it,” Willard said. “No one has figured out how to use the potential because it takes a lot of rethinking when writing software.”
Through its former owner, British-based International Computers Ltd., which spun off 80% of AMT in 1986, the Irvine firm has developed software based on a modified version of the Fortran programming language. Manning said the software gives the company a head start on would-be challengers.
Since 1987, AMT has sold only about 75 computers. Thinking Machines has sold 60 computers since 1983, while MasPar has sold seven since introducing its first line in January.
Manning’s biggest goal is to secure a significant sale to a defense customer or major oil company to help defray the costs of future product development.
But he worries that other larger companies such as International Business Machines Corp., which has invested in a small parallel computer firm founded by former Cray Research chief scientist Steve Chen, may muscle in on AMT’s territory.
