Hybrid architectures dominate Green500
- By Joab Jackson
- Nov 24, 2008
The top end of the fourth Top500
list of supercomputers, which suggests that specialized processors could lead to big power savings in supercomputing.
The latest iteration of the Green500 list was released at the SC08 supercomputing conference in Austin, Texas. It is the first time high-performance computers have executed more than 500 million floating-point operations/sec (megaflops) for every watt of energy they used, according to the list's compilers.
The most energy-efficient supercomputer is a 2,016-processor machine at the University of Warsaw's Interdisciplinary Centre for Mathematical and Computational Modeling
. The system, based on IBM BladeCenter QS22 servers, produced more than 536 megaflops per watt.
That system and the next six systems on the list run IBM's multicore Cell Broadband Engine (Cell/B.E.) processor. Four of those systems use IBM's new QS22 blade server. The top four machines all achieved more than 500 megaflops per watt.
Overall, the participating supercomputers showed a 17 percent increase in energy efficiency since the last edition of the ranking, released in June.
Wu Feng and Kirk Cameron, associate professors at Virginia Tech, started the Green500 in 2006 as a way to encourage supercomputer developers to think more about the amount of energy their creations consume. In the past few years, supercomputer managers have started to notice that the increasing size of their machines brought about a corresponding increase in energy bills.
"For decades now, the notion of performance has been synonymous with speed," the Green500 Web site states. 'This particular focus has led to the emergence of supercomputers that consume egregious amounts of electrical power and produce so much heat that extravagant cooling facilities must be constructed to ensure proper operation.'
The Green500 list is compiled from the biannual Top500 list of supercomputers, the latest iteration of which was released earlier this week
Top500 participants were encouraged to submit data on the average amount of energy their systems used during a Linpack run, a benchmark that determines their rankings on the Top500.
Positions on the Green500 are calculated by dividing the Linpack score, measured in flops, by the average amount of wattage used during the testing.
In the latest Green500 compilation, 276 of the top 500 sites submitted their energy usage, a 19.5 percent increase from the number of submissions in June. The list's compilers estimated the rest of the scores.
Comparing the Top500 with the Green500 list shows that the biggest computers are not always the most energy efficient.
The machine that topped the Top500 list ' Los Alamos National Laboratory's IBM-based Roadrunner ' ranked only ninth in the Green500. And the Top500 second place entry' Oak Ridge National Laboratory's Cray XT5 Jaguar ' ranked 80th on the green list, with 153 megaflops per watt. By contrast, the University of Warsaw system ranked 221 on the Top500 list.
The sudden influx of systems running specialized processors could point the way to more energy-efficient computing. Specialized processors, such as the Cell/B.E.
, Nvidia's Cuda
-based graphical processor units or plug-in field-programmable gate arrays, can provide more flops per energy consumed than general-use processors from Intel or Advanced Micro Devices.
Industry experts estimate that the Cell/B.E. can produce about 14 flops/sec for about 97 watts of energy, while a generic x86 processor can produce only about 1 flop/sec at that wattage.
The prominence of Cell/B.E.-based servers at the top of the lists suggests the growing influence of multicore processors, Cameron said.
Although IBM developed the Cell/B.E. with Sony for game-console use, it has expanded its applications for scientific computing. The chip has eight computation cores and one for coordinating traffic among the cores and between the chip and the rest of the system.
IBM offers a version of the technology in its QS22 blade server
. Mercury Computer Systems also offers a line of Cell/B.E.-based servers
Maciej Remiszewski, an IBM representative who worked with the University of Warsaw, said the center did not specifically configure the system for maximum energy efficiency, though obviously it was one of the benefits.
"The choice to go ahead with this architecture [answers the challenge] of how do you use the same amount of transistors and build a more computationally efficient system," Remiszewski said.
Maciej Filocha, an administrator of the Warsaw system, said the center will use the system to investigate better methods of programming Cell/B.E. processors.
Feng said the Green500 team will continue to analyze the results of the list and offer further insights in the weeks to come.