Virginia Tech College of Engineering
NATIONAL CAPITAL REGION, June 30, 2010 —With the latest release of The Green500 List, accelerator-based supercomputers now occupy the top eight slots of the Green500, where the 'fuel efficiency' (or energy efficiency) of supercomputers is defined as 'millions of floating-point operations per second' (MFLOPS) divided by 'watts' (W) or MFLOPS/W. Accelerators refer to the use of dedicated hardware to perform computations faster than a traditional processor, also known as a central processing unit (CPU).
Green500 co-founder Wu Feng, associate professor of computer science and electrical & computer engineering at the College of Engineering at Virginia Tech explained the significance of the 'fuel efficiency' of these accelerator-based supercomputers. “The accelerator-based supercomputers on The Green500 List produce an average efficiency of 554 MFLOPS/W whereas the other measured supercomputers on the list produce an average efficiency of 181 MFLOPS/W. That makes the accelerator-based supercomputers on the Green500 more than three times more energy efficient than their non-accelerated counterparts on the list” Feng said.
The accelerator-based supercomputers come in two flavors: one is based on the custom PowerXCell 8i processor from IBM; and the second is based on the commodity graphics processing unit (GPU) from one of two companies, either Advanced Micro Devices' (AMD) ATI technology or from NVIDIA. As in the previous edition of the list from November 2009, the former flavor tops the Green500 with three IBM quantum chromodynamics parallel computing on the cell (QPACE) machines, all tying for first place and all located in Germany at the University of Wuppertal, University of Regensburg, and Jülich Research Center, respectively. The low-power QPACE clusters use the IBM PowerXCell 8i processor, an enhancement of the Cell Broadband Engine originally developed by Sony, Toshiba, and IBM for Sony's PlayStation 3, as well as a network of programmable units called field programmable gate arrays or FPGAs.
In contrast, China has taken a notably different approach by leveraging the commodity graphics processing unit (GPU) as an accelerator. “Their first GPU-based supercomputer, Tianhe-1, debuted on the Green500's November 2009 list at #8 and used AMD/ATI Radeon HD 4870 GPUs. It has since dropped to #11. Two new GPU-based supercomputers, Dawning TC3600 and Mole-8.5, both from China, leapfrog Tianhe-1 and use NVIDIA C2050 GPUs to debut at #4 and #8, respectively,” Feng said. A GPU is traditionally used to accelerate the rendering of two-dimensional (2-D) or three-dimensional (3-D) graphics on a display monitor, possibly a computer laptop or a desktop display. For the GPU supercomputers from China, however, the GPUs are “re-purposed” to perform general-purpose computation on the GPU, i.e., GPGPU.
The Green500's exploratory Little Green500 List debuted in November 2009 with the intent to further raise awareness by driving energy efficiency as a first-order design constraint on par with performance, or more specifically, speed. For the Little Green500 List, the average efficiency is 199 MFLOPS/W, which is 18 MFLOPS/W higher than the average efficiency of The Green500 List. This efficiency improvement, however, comes at the expense of lower overall performance. “This exploratory list seeks to be more inclusive of the high-performance computing community by being more open in its definition of a supercomputer. The Little Green500 List ranks the energy efficiency of a larger set of supercomputers, where a supercomputer is defined as being as fast as the 500th-ranked supercomputer on the TOP500 List 18 months prior,” Feng said.
The Green500's exploratory HPCC Green500 List, where HPCC stands for the High-Performance Challenge Benchmarks, was announced in November 2009 and welcomes its first official entry, the Talon supercomputer, from Mississippi State University. “In addition to its listing in the HPCC Green500 List, Talon is currently ranked #9 on the Green500 and #331 on the TOP500.