RealTime IT News

Big Blue's Baby Blue Gene Supercomputer

You can call it a "Mini-Me" supercomputer but just don't call it Blue Gene Lite. IBM's researchers don't like that very much.

IBM unveiled a smaller version of its highly-touted Blue Gene supercomputer Friday just in time to crack into the coveted Top 500 Supercomputing list and days before the annual Supercomputing 2003 event in Phoenix, which begins Monday.

While the full Linux-based supercomputer won't be finished until 2005, the Armonk, N.Y. company has warmed up to the final bow by producing Blue Gene/L, a much smaller version of the project that is about the size of a 30-inch screen television. But, said IBM Deep Computing Vice President Dave Turek, don't be fooled by the petite nature of the product in a realm ruled by machines that fill whole rooms.

Blue Gene/L has been clocked at a peak speed of 2 teraflops, or 2 trillion mathematical operations per second, and currently sits at number 73 in the current Top 500 list, which will be formally issued at the event next week.

Turek said the computer is just a prototype at this point and is not geared for commercial rollout. But despite the fact that it only measures about one cubic meter, Turek said the machine stacks up favorably with most machines with comparable compute power, such as Linux clusters offered by Dell or SGI. Turek said most 2 teraflop supercomputers fill up entire rooms, often with more than a dozen racks.

"This is really a marriage of two ideas," Turek told internetnews.com. "Customers want amplified computer power but have legitimate concerns over space and cooling. We think the days of people building new buildings for computers have gone by the wayside. By progressively compressing computing in these form factors, we will be able to deliver supercomputing into the hands of customers such as small and medium-sized business who could not accommodate it today because they don't have the space for massive machines."

"What IBM has here is a research project for the Department of Energy that is set to scale to a massive size yet provide a compact footprint for low power consumption while providing high performance," DH Brown analyst Rich Partridge said.

Partridge said the news is proof that IBM has a physical product that works and has been recognized by the prestigious supercomputing list, bit cautioned that is not a commercial viability at this point. "This is not just a paper tiger -- it's a working, functional product."

So, could this mini Blue Gene/L parallel the advancement of thin, yet powerful form factors like blade servers, which have caught on with big systems vendors like IBM, Dell, HP and Sun Microsystems in recent years? That is, could IBM's work bring on a raft of smaller, modular supercomputers from the aforementioned competitors?

Turek said no, because supercomputing on its face is too complex for rivals to closely approximate what another rival is doing. There is too little commonality associated with the architectures, which would be a requirement for the commoditization of mini supercomputers.

"With the notion of commoditization comes the behavior of certain market dynamics," Turek said. "There is the fundamental economics of technology to consider but also the acceptance of how it's presented. If you decide you want to make toothpicks made out of thread instead of wood, they may be cheaper but the demand wouldn't necessarily drive the supply. It has to be something that people fundamentally want."

Moreover, Turek said IBM and rivals face the tough task of scaling up, which he said is never easy for supercomputers because there are a number of barriers associated with adding power or nodes to already massive computers.

"There comes a point in scaling up that you run into a wall if you haven't thought it through before hand," Turek said.

Turek said the original Blue Gene plan, when it was unveiled in 1999, called for a forest of computers in racks or cubes that could be linked together as sort of a "modular" supercomputer. The mini BlueGene/L is a scaled-down sample of what the ultimate machine will be like, which is being built for the Lawrence Livermore National Laboratory in California.

To date, Blue Gene/L has been used to map protein folding in the life sciences space, but Turek envisions it will have the potential to be used for any number of industries, including digital entertainment and financial services.

Partridge said it is possible Blue Gene/L could be commercialized for areas such as seismic engineering or fluid dynamics, and sell some units among the high performance technical computing arena, but wondered whether or not it wouldn't interview with IBM's current Linux supercluster products, powered by Opteron, Xeon and Itanium processors.

By 2005, Turek said Blue Gene/L will consist of a group of machines that will be 128 times larger than today's prototype and occupy 64 full racks. It is expected to operate at about 360 teraflops LLNL researchers hope to use Blue Gene/L to investigate areas such as cosmology and the behavior of stellar binary pairs, laser-plasma interactions, and the behavior of nuclear explosives.

To wit, Blue Gene/L is part of the National Nuclear Security Administration (NNSA)'s Advanced Simulation and Computing (ASC) Program.