Twice a year, in step with the biannual TOP500 list, the Green500 list ranks the most powerful systems in the world based on energy-efficiency. Published Wednesday evening at SC13, this year’s Green500 list continues a trend from previous years, the rise of heterogenous supercomputing. The latest list shows that the top 10 greenest systems are powered by NVIDIA Tesla GPUs, specifically the Kepler parts.
The only other architecture to have accomplished a clean sweep of the 10 top spots on the list is IBM’s BlueGene system. In fact, the top 20 spots on the June 2012 list were all occupied by IBM Blue Gene/Q supercomputers.
A heterogenous computing system employs two or more types of processing technologies, such as traditional processors (CPUs), graphics processing units (GPUs), and coprocessors.
The age of mixed-processor systems can be traced back to Roadrunner, which paired 12 thousand IBM PowerXCell 8i coprocessors with six thousand standard dual-core x86 CPUs. In June 2008, Roadrunner became the first supercomputer to deliver over a Linpack petaflop, earning it a number-one spot on the TOP500 list and a number-three spot on the Green500.
This year’s most energy-efficient supercomputer is the Tsubame-KFC system, installed at the Tokyo Institute of Technology. Boasting a record 4.5 gigaflops per watt, Tsubame-KFC is about 25 percent more efficient than the runner-up Cambridge University’s Wilkes supercomputer, which can process 3.6 gigaflops per watt. In third place is the HA-PACS TCA system at the University of Tsukuba, delivering 3.5 gigaflops per watt.
Many of the greenest systems of late have been relatively small, raising the question of whether energy-efficient techniques have scaling power. So it’s promising news that this latest top ten grouping includes two petaflop systems. Piz Daint, the 6.27 petaflop system at Swiss National Supercomputing Center, comes in at number four, delivering 3.2 gigaflops per watt, while TSUBAME 2.5, a 2.8 petaflopper at Tokyo Institute of Technology, sits in sixth position, with 3.07 gigaflops per watt.
The current fastest supercomputer, China’s Tianhe-2, relies on a heterogenous design for its record-breaking 33.86 Linpack petaflops. Equipped with Intel Xeon Phi coprocessors, Tianhe-2 achieves an efficiency of 1.9 gigaflops/watt, for a not-too-shabby number 40 Green500 ranking.
In an official blog post, a NVIDIA rep characterizes the company’s presence on the Green500 as ascendant. Six months ago, on the previous Green500 list, there were two systems in the top 10 with GPU parts.
“At the heart of this trend is the spread of NVIDIA Tesla GPU accelerators based on our Kepler architecture,” remarks the graphics chipmaker. “Launched last year, they are three times more energy efficient than the Fermi-based family of processors they succeeded.”
NVIDIA points to energy-efficiency as a crucial consideration for supercomputing systems going forward. The largest supercomputers require many megawatts of power and the cost can run into millions of dollars each year. Reaching exascale means boosting speeds by 50-100 times, while keeping power relatively static.
These power constraints need to be addressed for exascale computing to be a reality. In a talk at SC13, the University of Tennessee’s Jack Dongarra said that achieving this feat will require a tenfold increase in efficiency, i.e., systems capable of 50 gigaflops per watt.
It’s a challenge that can sound daunting, but Green500 figures show a better than tenfold gain in efficiency over the last six years. On the November 2007 list, England’s Daresbury Laboratory system, an 11.1 teraflops (Rmax) Blue Gene/P, operated at 0.3 gigaflops a watt.