Here is a recent article that shows how supercomputing is going at http://hardware.silicon.com/storage/0,39024649,39127800,00.htm

This is not one of the elite top-500 supercomputers.

IBM to sell supercomputing server

February 11 2005

by Stephen Shankland

Taking high-performance to the next level?

IBM plans to begin selling on 18 February a powerful but power-hungry server specialised for supercomputing tasks, silicon.com's sister site CNET News.com has learned.

The p5-575 combines eight Power5 processors and as much as 32GB of memory into a computing slab 8.9cm tall, 61cm wide and 127cm deep. IBM will begin selling the servers in clusters of eight on 18 February and clusters of 128 on 29 April, according to IBM's website.

IBM plans to show off the systems next week at the LinuxWorld Conference and Expo in Boston. It already announced many system details at a November supercomputing show.

One of IBM's top server priorities today is high-performance technical computing. The company already has a multitude of products for the market, ranging from ordinary clusters of lower-end servers to the exotic Blue Gene/L.

Illuminata analyst Gordon Haff said: "High-performance computing is strategic for everyone, because it's becoming more and more mainstream." For one thing, more customers are performing technical computing tasks, such as mining customer data for purchasing and supply trends. For another, general-purpose servers are getting better at doing the technical computing work.

The p5-575 machines can run AIX, or Linux from Red Hat or Novell. A single node, including AIX, 16GB of memory and a year of maintenance, costs $92,000, IBM said.

The systems are massive and power-hungry.

A single rack, topped up with the maximum of 12 systems, power systems, an input-output module and noise-dampening doors, weighs 3,460 pounds. And those 12 computing nodes consume up to 41.6kW.

"With the density levels you're starting to get today with blades and servers like the 575, fairly specialised power and cooling is once again coming to the forefront" of data centre operations, Haff said. Systems such as these typically require specially reinforced floors, he said.

Each system uses eight 1.9GHz Power5 processors. Ordinary Power5 chips have dual processing engines, called cores, but the p5-575 chips have only one core active. That means more high-speed cache memory and communication bandwidth is available for that core.

Most supercomputing clusters use lower-end dual-processor servers with Intel's Xeon. However, customers often need larger amounts of memory that those machines can't accommodate.

That was the case for the p5-575's star customer, Lawrence Livermore National Laboratory's ASC Purple, a supercomputer for predicting whether nuclear weapons will continue to work as they age.

The lab needed systems with large amounts of memory for ready access to tables of data on which calculations are based, the lab has said.

Initially, the Livermore lab had planned to use higher-end systems with as many as 32 dual-core Power5 processors, but a budget cut led the lab to opt instead for the p5-575. Changing the technology saved money but resulted in a six-month project delay, a representative said.

The lab is consuming much of the p5-575 production capacity, but at least three other customers will have systems by the end of the quarter, IBM said.

IBM sells its own switch to link computers into a high-performance computing cluster. That switch won't support the p5-575 until the second quarter, IBM said.

Stephen Shankland writes for CNET News.com.