Lights out for original 'fastest super'

The lights will go out on a piece of computing history May 29 when the California Institute of Technology pulls the plug on the first scalable parallel supercomputer to earn the title of fastest in the world. Built by Intel Corp. as a prototype seven years ago, mostly with federal funds, the 512pr

The lights will go out on a piece of computing history May 29 when the California Institute of Technology pulls the plug on the first scalable parallel supercomputer to earn the title of fastest in the world.

Built by Intel Corp. as a prototype seven years ago, mostly with federal funds, the 512-processor Touchstone Delta spawned a new era of scientific computing. The system gave government and university researchers, for the first time, enough computing power to pursue such groundbreaking applications as real-time processing of satellite images and molecular models for AIDS research.

Although a second version was never built, the Delta was the foundation for even faster computers— some with thousands of processors— made by Intel and others and installed today at NASA, the Defense Department and the Energy Department as well as other federal labs. Those include systems made by Intel, IBM Corp. and Silicon Graphics Inc./Cray Research that are being used to model the condition of the nation's nuclear stockpile under DOE's Accelerated Strategic Computing Initiative (ASCI).

"We are all so accepting of so many things today,'' said Gilbert Weigand, who is now DOE deputy assistant secretary for strategic computing and simulation, but was then the Defense Advanced Research Projects Agency program manager who gave Delta the go-ahead. "At that time, people really weren't sure it would work. There were people even four years earlier that didn't believe you could use a computer with 250 processors.''

Meanwhile, the cooperation among DARPA and other federal agencies to fund the $18 million project helped forge partnerships that continue today in the multiple-agency Computing, Information and Communications program, the government's $1 billion high-performance computing and network research initiative. "We had to find a way for everyone to share this large resource,'' Weigand said. "Everyone got early results that helped bolster the [high-performance computing] program.''

Intel never planned to build a prototype as large as the Delta, said Justin

Rattner, who designed the system and is now the director of Intel's Beaverton, Ore., Server Architecture Laboratory. However, Rattner said, "it had always been our experience that whenever we proposed to build a machine of a given size, the potential users wanted a machine at least twice that big.''

"We believed that industry had learned how to build much more powerful parallel computer systems, ones that would rival'' the vector-based systems that predominated, recalled Paul Messina, director of the Center for Advanced Computing Research at Caltech. Caltech scientists researching biology, chemistry, astronomy and aeronautics had applications to run on such a machine, and they wanted the chance to use one.

Messina and his colleagues began knocking on doors in Washington, D.C., until they pulled together enough grant money to fund the system. The resulting Concurrent Supercomputing Consortium, which included Caltech, DARPA, DOE, NASA, the National Science Foundation, Rice University and Purdue University, chose Intel to develop it.

"Our plan worked pretty well,'' Messina said. "The Delta was the world's most powerful scientific computer for about 18 months, and many important calculations were carried out on it.'' Among its achievements, the system, which reached a peak performance of 32 billion floating-point operations per second, was used by NASA to process 3-D images of Venus taken by the Magellan spacecraft during the early 1990s in near real time.

Intel stopped producing Delta's descendant, the Paragon, after delivering a version 30 times more powerful to Sandia National Laboratories for the ASCI program last year.

Meanwhile, DOE has funded IBM, SGI/Cray, Digital Equipment Corp. and Sun Microsystems Inc. to create a computer 3,000 times faster than the Delta within the next six years.

This computer will be built from "commodity'' chips and other general-purpose parts— another legacy of the Delta experiment. "What the machine did was it showed that you could take commodity parts with an interconnect and build a large supercomputer,'' said Horst Simon, director of the DOE National Energy Research Supercomputer Center.

Delta still works; its nine black cabinets flash green as its processors fire. But with more powerful machines running alongside it, it is mainly valuable today as a museum piece. Federal users ported their applications to other systems a year ago.