Big Blue captures three victories in the supercomputer realm and wins two major new contracts for nuclear weapons and climate research.
In a research project, IBM will build a system called Blue Gene/L with 65,000 processors at Lawrence Livermore National Laboratory (LLNL) for nuclear weapons simulation, the company plans to announce Friday. The machine is an extension of the Accelerated Strategic Computing Initiative that has paid for the top-ranked machine, IBM's ASCI White, also at LLNL.
Blue Gene/L will have 16 trillion bytes of memory--about 100,000 times the amount of a typical PC today--and will be able to perform 200 trillion calculations per second, the same rate as each person on the planet punching 33,000 calculations a second on a hand calculator.
The machine is part of IBM's $100 million "Blue Gene" program to build a computer by 2005 or early 2006 that can perform a quadrillion calculations per second, said William Pulleyblank, director of Exploratory Server Systems in IBM's Research division. Blue Gene/L, due in 2004 or 2005, will have about one-fifth that horsepower but will be able to transfer data faster.
IBM also won a $24 million contract to build "Blue Sky" at the National Center for Atmospheric Research (NCAR) by September 2002. The system will be used to predict climate changes such as global warming, a computationally intensive task that thus far hasn't satisfied scientists' demands for detailed projections.
Supercomputers such as these have largely been a subject of interest to government and academic researchers, but businesses also are buying the systems for tasks such as modeling a car's response in a crash, the aerodynamics of a jet or patterns of credit card fraud.
Supercomputers carry more than big price tags. Supercomputer efforts indicate the depth of research on a product line that companies often commit to for years.
In the independent and twice-annual Top 500 supercomputer ranking that appears on the eve of the SC2001 supercomputing show, Hewlett-Packard and Compaq Computer stole some spots from IBM. But Big Blue maintained the top ranking it definitively won last year with the fastest system, five of the top 10 systems, 32 percent of systems on the entire list and 37 percent of the combined computing horsepower.
The Superdome system gave HP a bigger presence on the previous Top 500 list in June, increasing from 41 systems on the list to 153, but most of those appear in the bottom 100. Compaq, meanwhile, has the No. 2 and No. 6 spots, but only has 16 total on the list.
SGI took third place with 41 systems, with Cray in fourth with 39 and Sun Microsystems in fifth with 31.
The Top 500 list is compiled by researchers from the University of Tennessee, Lawrence Berkeley National Laboratory and the University of Mannheim. The researchers are working with research firm IDC to create a benchmark that better corresponds to real-world computing performance, however.
Blue Gene/L is named after IBM's "L" architecture, a different design from the "C" design that underlies the earlier-announced Blue Gene machine, which will be used to figure out the "folding" process by which genetic instructions manufacture protein molecules.
The C architecture would use a large number of chips--about a million CPUs all together, grouped 32 or so apiece on a single slab of silicon. The L architecture uses chips more similar to IBM's Power line of CPUs, and uses fewer of them, Pulleyblank said.
The system will have about 65,000 two-CPU modules, Pulleyblank said. They'll be stacked 1,000 at a time into refrigerator-sized cabinets that will together occupy about half the size of a tennis court--considerably smaller than ASCI White and a more powerful Compaq ASCI machine now under construction at Los Alamos National Laboratory.
In Blue Gene/L, each processor in the two-processor nodes will have a separate task. One will handle numerical calculations, while the other will handle communication with the CPUs across the rest of the system, Pulleyblank said.
One advantage of Blue Gene/L is that the central control center can broadcast messages simultaneously to all nodes on the system or harvest data likewise in one fell swoop, he said. That's much faster than having to poll each processor sequentially.
Each cabinet has slots that can accommodate boards either with CPU modules or input-output modules for communication across the overall system, Pulleyblank said.