The next-generation prototype of IBM's Blue Gene, Blue Gene/Q, has topped the latest iteration of the Green500, a ranking of supercomputers by their power efficiency, released today at the SC2010 conference in New Orleans.
Blue Gene/Q is 165% more efficient than the Chinese supercomputer, Tianhe-1A, that topped the latest Top500 list, released Sunday. It is 77% more power-efficient than the number-two entry on the Green500 list, the Tokyo Institute of Technology's Tsubame 2.0.
In terms of raw numbers, Blue Gene/Q was shown to execute 1,684 megaflops per watt. Tsubame 2.0 demonstrated 948 megaflops per watt, and the EcoG system from the U.S. National Science Foundation's National Center for Supercomputing Applications placed third with 933 megaflops per watt.
The list shows that the most powerful supercomputers may not be the world's most efficient users of energy. Blue Gene/Q, for example, demonstrated a peak performance of only 653 teraflops, a small number compared with Tianhe-1A's 2.57 petaflops. However, Tianhe-1A only demonstrated an efficiency of 635 megaflops per watt, and placed tenth on the Green500.
Virginia Tech researcher Wu-chun Feng created the Green500 in 2007 as a way to draw attention to the growing energy consumption in supercomputers. "We consider [energy use] to be a first-order design constraint," Feng said in an interview with the IDG News Service.
Feng noted that by the end of the next decade, if system designs continue on their current trajectory, the cutting-edge supercomputer will consume a gigawatt, or 1 billion watts, of power. In contrast, the entire state of New York consumes about an average of 62 gigawatts per year.
The list is updated twice a year. Participation is voluntary, and rankings are based on how many floating-point operations are executed per watt of electricity. Only those computers placing in the Top500 are considered for the list.
This latest iteration of the test shows the growing use of GPUs (graphics processing units) as a relatively low-power way to add more computing power. Four of the top 10 entries on the list, for example, use Nvidia GPUs, which would have been unheard of a few years ago.
IBM's Blue Gene/Q did not incorporate GPUs but rather relied on another type of accelerator card, the IBM PowerXCell-based accelerator.
"2010 could be seen as being the year of the accelerator," Feng said. "It's an alternative computing platform that maps well to various tasks. The side benefit, which is very important, is that they happen to be energy-efficient."