Big Blue will build a machine at Los Alamos lab that combines thousands of AMD Opteron and Cell chips.
As first reported by CNET News.com, the machine, dubbed Roadrunner, uses a hybrid approach that combines a conventional cluster of Opteron servers with Cell chips that handle some of the calculating grunt work. Each Cell chip, originally designed by IBM, Sony and Toshiba for the Sony PlayStation 3 video game console, includes eight special-purpose engines that can rapidly perform physics calculations.
IBM and the Energy Department's National Nuclear Security Administration announced Wednesday that Big Blue had won the contract. Pete Domenici, a Republican senator from New Mexico, where the nuclear weapons lab is located, said of the deal, "It's time to restore LANL to the forefront of computing technology. Together with IBM, the lab will undertake an exciting goal of creating the world's fastest supercomputer."
LANL's sister lab, Lawrence Livermore National Laboratory, currently houses the top-ranked machine, IBM's Blue Gene/L, which can perform 280 trillion calculations per second, or 280 teraflops. Roadrunner is designed to nearly quadruple that to a sustained speed of 1 quadrillion floating-point operations per second, or a petaflop.
Roadrunner, which will run Linux and include software to juggle tasks between the Opteron and Cell processors, will be built using commercially available IBM hardware. That includes System x3755 servers with four Opteron processors apiece and IBM BladeCenter H servers with Cell-based systems.
IBM and LANL aren't the only outfits gunning for petaflop supercomputers.
Supercomputer specialist Cray plans to build a petaflop machine for Oak Ridge National Laboratory, and won a contract to build a machine for the National Energy Research Scientific Computing Center (NERSC) that has an option to expand to a petaflop level.
And Japan's Institute of Physical and Chemical Research, called RIKEN, said in June its Protein Explorer supercomputer already has reached the petaflop level, though not using the conventional Linpack supercomputing speed test.