Livermore's IBM BlueGene/P will be chased by one from Europe
If the US government thought that investing DOE money into a 20 petaflop computer would give the country a competitive advantage, it learned today it's mistaken: IBM will also be partnering with German researchers.
While the initial goal for the BlueGene/P model being developed for the Forschungszentrum Juelich's Gauss Centre for Supercomputing will be to break just the one petaflop barrier -- the one thousand trillion floating-point instruction mark already superseded by two US Dept. of Energy supercomputers with IBM's help -- the design chosen is the same one the company announced last week for Lawrence Livermore Labs.
In its initial incarnation at least, the Gauss BlueGene/P won't have the processor power to aim for the 20 petaflop model that IBM and the DoE outlined last week. But with a respectable 294,912 PowerPC 450 processor cores at 850 MHz (not necessarily the top of the line, especially at 32-bit), it would appear on the surface that it should have plenty of muscle to edge out Livermore's former champion BlueGene/L. A BlueGene/P design with 163,840 PowerPC cores with the same buildout, named Intrepid, placed fifth on last November's Top 500 Supercomputers list. Maintained by Argonne National Laboratory, a DoE research facility in Illinois, it presented an Rmax score of 450.3.
All things being equal, just adding more processors to the mix would give the Gauss system a theoretical projected score of 810.54, well below the petaflop barrier. If it's going to beat Oak Ridge Labs' Cray Jaguar -- the current #2, and the second system to cross the petaflop mark -- the Gauss BlueGene/P will have to innovate in some different departments. Gauss hopes an innovative water-cooling system, leading to a reduction of 91% of the system's necessary air conditioning, may lead to performance improvements.
But being cool and clammy doesn't always pave the way for speed. A 2008 evaluation of BlueGene/P architecture (PDF available here) by Oak Ridge -- the one running Jaguar that beat Intrepid -- concluded that while the PowerPC processors behaved well in the power reduction department, that didn't mean they performed when it counts. "In summary, BG/P performs very well on power metrics across the board; however, its advantages are much less when considering science-driven workloads, like POP [Parallel Ocean Program, an ocean wave simulator model], and taking into account the aggregate amount of power necessary to obtain specific levels of computation throughput."