Some of the most powerful supercomputers in the world are also the most power efficient, according to Erich Strohmaier, head of the Future Technology Group at the Lawrence Berkeley National Laboratory (LBNL).
The average power consumption of a TOP10 system is 4.09 MW (down from 4.56MW six months ago). By comparison, the average power consumption of a TOP500 system is 671.3kW (up from 634kW six months ago and 543kW one year ago).
Speaking at the International Supercomputing Conference in Hamburg, Strohmaier said that in November 2011, many of the TOP10 supercomputers were using old technology that was less power efficient. Many of these old systems have now been replaced by new systems, bringing the average power consumption down.
“As an average, the TOP50 was actually more up-to-date than the TOP10 six months ago, but that has changed. The TOP10 now has the new technology, which shows that supercomputing is very much a market where innovation comes from the top and trickles down,” said Strohmaier.
“The new stuff, as soon as we have confidence that it works, it’s going to get introduced in a very large system. That’s why the large systems have higher efficiency than the small systems.”
Overall, only 40 systems on the list are confirmed to use more than 1MW of power. The number one system, Sequoia, an IBM Bluegene/Q system, reports the second highest total power consumption of 7.89MW, but is deemed to be one of the most energy efficient systems on the list.
Strohmaier said that, by the end of the decade, all of the top 10 systems will consume around 8-9MW of power – a significant step down from LBNL’s prediction six months ago of 13-14MW. For this to happen, however, there will need to be dramatic improvements in technology at regular intervals.
“We can’t just sit back and say everything’s good, we’re going to be fine in 10 years. That’s not the case. There is a lot of work needed to arrive there,” he said.
The power consumption of supercomputers is one of the major issues facing the industry, because as clusters get bigger and more powerful, there is often not enough energy to keep them running. At the European Nuclear Research Organisation (CERN) near Geneva, for example, scientists often complain that power limitations are having a knock-on effect on the performance of their experiments.
Many chip manufacturers are now working to tackle this issue. Today’s most energy-efficient supercomputers include:
- An Intel cluster with SandyBridge and MIC technology installed at Intel with 1176 MFlops/watt (number 150 on the list)
- An Appro Xtreme-X SandyBridge-based cluster installed at Los Alamos National Laboratory with 1050 MFlops/watt (number 73)
- A Chinese system called Mole-8.5 with nVidia accelerators with 919 Mflops/watt (number 21)
- The Fujitsu K Computer at RIKEN with 830 Mflops/watt (number 2)