SGI/Cray system lowers vector computing cost

Silicon Graphics Inc./Cray Research last week unveiled a new supercomputer based on its popular vector technology that is designed to provide users of its earlier, proprietary Cray computers with a migration path toward future systems built from commodity parts.

The Cray Scalable Vector 1 (SV1), introduced June 16 at a Cray User Group meeting in Stuttgart, Germany, represents the first step on a product development path that SGI presented this spring to merge its technology with that acquired from Cray and migrate to a platform based on Intel Corp. processors.

Many supercomputer users are migrating from vector-based processing to scalable parallel computing platforms, including other Cray-designed systems. But some customers wish to remain on vector technology.

"The SV1 is designed for those customers and users who see benefits in continued vector [processing]," said Bill Minto, director of platform product management at Cray. "It remains an important market for us from a revenue perspective, and for the customers it enables them to develop next-generation applications."

It has been more than two years since Cray upgraded its current J90 supercomputer. Some industry analysts observed that the company has failed to keep up with industrywide price/performance improvements, including products from major Japanese competitors, although federal customers have remained loyal to Cray as the only U.S.-based vector supercomputer maker.

While SGI/Cray redesigns its product line, "in the interim, what they're saying is, 'I'm going to lower the cost and bring it much more in line,' " said Richard Partridge, a vice president with technology evaluation firm D.H. Brown Associates Inc.

The SV1, which includes faster processors and a new vector processor cache, would match the performance of the J90 at one-fifth the cost. The technology also is designed to allow users to build very large systems that are capable of 1 trillion floating point operations per second (teraflops), the company said.

Brett Berlin, a consultant who advises the Defense Department on high-performance computing, said the announcement is good news for federal users who still want to use the traditional vector supercomputing technology. Competing scalable parallel computer designs "have still not delivered for a lot of the very large vector applications" such as running climate models and analyzing intelligence, he said.

"It's almost a life-support system for some of these customers," said Gary Smaby, president of Smaby Group, a market research firm. "Anybody that was committed to a classic vector architecture had been expecting these midlife kickers," he said, although he doubted the machine would be appealing to anyone developing new applications.

In fact, large-scale federal supercomputer users have been steadily migrating to scalable parallel machines, which differ from vector systems in the way they perform calculations. Those users have found it less expensive to build very large systems using a scalable parallel design, partly because these systems are based on widely manufactured components that also are used in servers and workstations.

But in some cases, Partridge said, agencies have not thought it cost-effective to spend money rewriting their software for new machines.

"Very few people at the high end have totally moved all their stuff off" vector systems, Berlin said.

Minto confirmed that the company is negotiating sales with federal customers, but he said he could not discuss any upcoming deals. The SV1 is slated to be available in August.

According to an analysis by D.H. Brown Associates, a 32-processor configuration would cost $5 million to $12.5 million. Users would require 32 such nodes to achieve the peak 1-teraflops performance.