The XD1 models, which are based on technology fromthis year, cost as little as $50,000 for a single-chassis system with 12 Advanced Micro Devices' Opteron processors, Cray said. Most customers are expected to buy three-chassis systems for about $100,000, though more powerful models will sell for about $2 million.
Cray specializes in high-performance technical computers, but faces competition from mainstream server companies adapting their products for the small but prestigious market. Rivals includeand Hewlett-Packard, the , as well as Dell, Silicon Graphics, Sun Microsystems and NEC.
Cray's new systems are designed to compete with a newer breed of high-performance computer: clusters of low-end servers, networked together and typically running Linux on Intel or AMD chips. These clusters often use a software layer called (MPI) that lets programs share data among different computers in the cluster.
"The Cray XD1 system promises to substantially outperform conventional Linux clusters in the same price range," the company said in a statement.
The XD1 systems use Linux and MPI, but replace the commonly used Ethernet network with a very high-speed interconnect that directly links processors and memory with the Opteron HyperTransport technology, Cray said.
"It provides very high bandwidth between the nodes and very little latency," or delay, said Steve Scott, a Cray chief architect.
Cray also is working on a high-end system called, whose design, including customized high-speed communication chips, parallels the XD1 on a larger scale. Cray is working on a unified successor to Red Storm and the XD1 code-named Mount Adams, a system that will pair the Opteron with a new communication chip drawing on the ones used in both Red Storm and the XD1, Scott said.
Mount Adams also will use the Opteron, he said, praising the open-standard HyperTransport link it employs. "They've just got this really great spigot you can connect to and build really great machines," he said.
Intel's front-side bus technology, by comparison, is "not open and does not give you the kind of direct high-bandwidth, low-latency access to local memory you get with a HyperTransport link," Scott said.