updated 06:30 pm EST, Wed November 23, 2011
Tsubame 2 supercomputer 3.4X more efficient
NVIDIA championed a milestone for supercomputers Wednesday with the launch of a new supercomputer cluster that focuses on eco-friendly power. Tsubame 2.0, at the Tokyo Institute of Technology, uses a mix of traditional processors with Tesla cards to dramatically step up the potential maximum performance without having to ramp up the energy it uses. Although it can calculate at up to 1.19 petaflops, it uses a comparatively modest 1.2MW of power, making it 3.4 times more efficient at getting work done than Los Alamos National Laboratory's processor-only Cielo Cray.
It's efficient enough that it ranks as the most efficient supercomputer on the Green500 to crack the petaflop level.
Each individual element of the new Tsubame cluster is an HP ProLiant SL390 server that has a strong Xeon chip behind it, but also with Tesla boards inside. The Tesla hardware's speed in certain kinds of general-purpose and often very parallel computing, such as vector math, lets it handle about 80 percent of the workload without needing either faster Xeons or more PCs.
AMD has general-purpose computing through its FireStream boards, but these have so far had trouble getting adoption at the supercomputer level. NVIDIA was the first to make cards dedicated solely to computing, where AMD for awhile was still depending on video cards using the extra computing power as a bonus.