SGI's Origin 3000 and Onyx 3000 servers are the first to offer "build to suit" configurations, the company said. "The concept of self-contained modules is new," said Jeffrey J. Hewitt, principal analyst for servers worldwide at Dataquest Inc. (San Jose, Calif.). "It's not unusual for supercomputers to offer nodes that can be tied together to create a larger system. But the idea of breaking the nodes into separate modules hasn't been done before."

Brick-based systems

SGI's modular approach, called NUMAflex, employs seven basic technology "bricks." The 19-inch-wide rack-mounted bricks include: a C-brick incorporating a CPU and memory; an I-brick that acts as a base I/O module; an R-brick with router interconnects; a P-brick with PCI expansion hardware; an X-brick with XIO expansion; a D-brick with disk expansion; and a G-brick with graphics extensions. All of the bricks plug into a common backplane that forms the backbone of the server.

Separation of the various modules enables users to tailor a machine to their application. For traditional number crunching applications, such as computational fluid dynamics or weather modeling, users can configure the system with a predominance of C-bricks, with CPUs and memory. For visualization applications, they can apply graphics bricks, and for signal processing, they can use PCI bricks. "If you need a lot of computing horsepower, you can add CPUs without having to add I/O that you don't need," Hewitt said. "Or if you need more I/O, you don't need to add extra CPUs."

The modular configuration enables SGI to offer the new servers at costs far below those of supercomputers, the company said. "This is different from supercomputing because you can start at the very low end of the cost scale," said Ben Passarelli, director of advanced system product marketing for SGI (Mountain View, Calif.). The new server lines start below $20,000, Passarelli said.

The starting cost places the new system at the high end of the server market, which is still much lower than the cost of a supercomputer.

SGI engineers said a key to the system is its use of "cascading switches" in the router interconnects. The cascading switches enable a system to exhibit low latencies, or memory delays, despite the modular design. That is critical in machines that use non-uniform memory architectures (NUMA), the company said. Latencies have always been an issue with NUMA architectures, the company said, because memories are distributed rather than contained in one location.

The cascading switches, which are essentially ASICs, use a directory-based memory scheme to keep track of the data contained in various caches. As a result, there's less data traffic moving back and forth between memories, which results in lower latencies than conventional bus-based systems. SGI said that the Onyx and Origin systems have worst-case latencies of about 200 nanoseconds, even for eight-processor systems. Some competing systems still measure their latencies in microseconds, SGI said.

Low latencies are critical to certain users, especially those involving heavy number crunching. "Those latencies might seem like small amounts of time individually, but in those compute-intensive applications, the latencies can build up," Hewitt said. The latencies are particularly important in applications that are broken into pieces and done in parallel by various processors, he said.

The real test for the new systems, however, will come when users put them to the test. "Less latency can look great on paper," Hewitt said. "But SGI will have to prove their worth in real applications. That's the only real accurate way to measure a system's performance."

Prior to Tuesday's announcement, SGI said it had already booked over $100 million worth of orders for the new systems, including 18 systems of 128 CPUs or greater.