U.S. supercomputing efforts are making up for the country’s earlier loss of the speed title tothe Japanese Earth Simulator, with two new systems poised togarner top spots on the Top500 Supercomputer list due out next week.
The U.S. Department of Energy this week announced that the IBM BlueGene/Lsystem at Lawrence Livermore National Laboratory achievedrecord-breaking performance of 70.72 trillion floating point operations persecond (teraflops), almost double the speed of Japan’s Earth Simulator, made by NEC, which has led the Top500 list for two years with a top speed of 35.86teraflops.
The BlueGene/L news camejust one week after NASA announced its SGI Columbia system, which achieved apreviously unprecedented speed of 42.7 teraflops.
One-Two Super Punch
Erich Strohmaier, co-compiler and editor of the Top500 Supercomputer Sitelist, told TechNewsWorld that the NEC Earth Simulator had taken U.S.supercomputer scientists by surprise, but that with advancing technology — in particular,more and faster processors and more efficient interconnection –the U.S. had made up the difference.
“The Earth Simulator announced three years ago was a bit of a shock,”Strohmaier said. “It took U.S. researchers awhile to catch up, but with theuse of different technologies, they have caught up.”
While he would not discuss who would hold the top spot on the new list, Strohmaier said that both the BlueGene/L and the Columbia system — already installed and in use atthe NASA Ames Research Center — had established claims for top spotson the list, which will be announced early next week.
Strohmaier also indicated that the big boosts in speed and performanceamong supercomputers is likely to continue.
“We see exponential growth a little faster than Moore’s Law,” he saidreferring to the Intel founder’s formula for doubled processor performanceevery 18 months. “It’s possible because these systems use more processorsalong with faster processors.”
In announcing the latest BlueGene/L capabilities, U.S. Secretary of EnergySpencer Abraham said in a statement that the system would reduce thetime-to-solution for complex problems including nuclear weapons analysis andapplied scientific research.
Tilak Agerwala, IBM Research’s vice president of systems, called BlueGene’s70 teraflop performance “a milestone.” He highlighted the collaboration withLLNL and the research possibilities of the system.
“It is very significant that BlueGene has achieved this big number,”Agerwala told TechNewsWorld.
He also indicated the supercomputer “leapfrogging” will likely continue. IBM, he said, is hoping to achieve a 300 teraflop system with LLNLnext year.
Both the BlueGene/L announcement this week and the Columbia announcementlast week centered on supercomputer systems that had not yet been put totheir full potential use.
Agerwala indicated that the BlueGene/L had achieved the 70 teraflop markusing only about one-quarter of the final machine.
Columbia — a cluster of 20 SGI machines with 10,240 Intel Itanium 2processors — achieved its speed of 42.7 teraflops using only 16 of thecluster’s 20 installed systems.
Experts indicated they expect the progress in supercomputing togo on, even in the face of increasing challenges on matters of heat, power,manageability and other factors.
“We see it continuing,” Strohmaier said, referring to programs such asDARPA’s High Productivity Computing Systems program. “Their goal is tobasically push the speed of performance at the rate it has been in thepast.”
IBM’s Agerwala said that although the growth of processor performance andfrequency is slowing down — mainly because of power constraints –researchers are still pushing ahead.
“I expect at the system level, performance will continue to growrapidly,” Agerwala said. “You just have to do careful integration betweenthe silicon devices and the system you support.”
Agerwala said the use of low-energy, low-power processors andsystem-on-a-chip technology in BlueGene/L, for example, would allow it tocontinue gaining in speed.
“We’ve integrated it all together in innovative ways to get a balancedsystem design,” he said. “Then we can scale on the large problems. I seeBlueGene continuing to scale effectively on some very interesting problems.”