Much of the attention at the SC14 supercomputing show last week in New Orleans centered on the release of the Top500 list, a roster of the world's fastest supercomputers. The twice-yearly list is leveraged by system vendors and component makers to tout their technologies and by countries for bragging rights.
However, in the computing world—not only in the high-performance computing (HPC) space, but also among enterprises, telecommunications providers and businesses in general—power efficiency often is as important a metric when assessing data center technologies as performance is. That is the driver behind another list that is unveiled during supercomputing shows twice a year—the Green500, a list of the world's most efficient systems.
In the most recent list unveiled Nov. 20 at the show in New Orleans, the L-CSC—housed at the GSI Helmholtz Center in Darmstadt, Germany—ascended to the top of the list, with a power efficiency of 5.27 gigaflops per watt (billions of operations per second per watt). The system was built by Asus and is powered by 10-core Intel Xeon E5-2690 v2 "Ivy Bridge" processors and FirePro S9150 GPU accelerators from Advanced Micro Devices.
GPU accelerators and x86 coprocessors played key roles in the development of the most efficient systems. Such accelerators are increasingly being used in general in HPC environments, with 75 of the systems on the Top500 list—and four of the top 100—running either GPUs from Nvidia or AMD or Xeon Phi coprocessors from Intel. For the Green500 list, the top 23 supercomputers leveraged accelerators, with the bulk being Nvidia GPUs. However, AMD GPUs and Intel Xeon Phis also were used in some of the systems.
In addition, the second most efficient supercomputer—the Suiren ExaScaler 32U256SC cluster, powered by 10-core Xeon E5-2660 v2 chips—used the many-core Pezy-SC processor from Japanese vendor Pezy Computing as custom accelerators. Each of the accelerators contain 1,024 processing elements that use a pair of ARM926 cores as controller cores, according to organizers of the Green500 list. The Pezy supercomputer, which is running at the KEK High Energy Accelerator Research Organization in Japan, came in with a power efficiency of almost 4.95 gigaflops per watt. It represents the first time a system appeared on the list that used accelerators from vendors other than Nvidia, AMD or Intel.
This latest list marks the first time an AMD accelerator was used in a system that topped the roster.
"This unique position can only be achieved through sustained innovation at the leading edge of the computing world and processor and system design," David Cummings, senior director and general manager of professional graphics at AMD, said in a statement.
Of the top 23 systems, 17 used GPU accelerators from Nvidia. Two used accelerators from AMD, and another three leveraged x86-based Xeon Phi coprocessors from Intel. The third most efficient system on the list—the Tsubame-KFC supercomputer developed by Hewlett-Packard and NEC—uses six-core Xeon Ef-2620v2 chips and Nvidia's Tesla K20x accelerators. It's housed at the GSIC Center at the Tokyo Institute of Technology in Japan.
The Tsubame system had been at the top of the Green500 list for the previous two editions before being overtaken by the L-CSC and Suiren supercomputers, according to list organizers. It has a power efficiency of more than 4.44 gigaflops per watt, and was the first system in the world to surpass 4 gigaflops per second.
The Piz Daint supercomputer, a Cray XC30 system powered by eight-core Xeon E5-2670 chips and Nvidia K20x GPU accelerators and housed at the Swiss National Supercomputing Center, was ranked on the Top500 as the world's sixth-fastest system and on the Green500 as the ninth most efficient.