In 2003, the US Department of Energy's Pacific Northwest National Laboratory (PNNL) installed a US$25 million supercomputing system with about 2,000 of Intel's Itanium 2 processors. At the time, the machine was recognized as the world's fifth most powerful computer.
That was then. Four years later, the PNNL's system has dropped to 113th place on the Top500, a list of supercomputers maintained by a group of university researchers. The Itanium-based hardware is no longer meeting the laboratory's computational needs and is being replaced by a more powerful system under a deal with Hewlett-Packard Co. that was announced on Friday.
HP, which also supplied the existing system to PNNL, said that the new one will contain 4,620 of Advanced Micro Devices' new quad-core Opteron processors. Peak performance will be about 163 TFLOPS, vs. 11.8 TFLOPS on the system now in use at the laboratory, according to HP.
The new system will cost the PNNL about US$24 million and is expected to be fully operational by next September. When the machine is put into use, HP believes it will rank among the top five supercomputers in the world in terms of performance.
The Itanium-based system will be retired, and even the new one has a fixed life expectancy. "My target now is to have a new replacement by the end of the fourth year," said Kevin Regimbal, manager of systems operations at the PNNL's Environmental Molecular Sciences Laboratory.
Although Regimbal is aware of where the PNNL's existing supercomputer stands on the Top500 list, regaining a high ranking isn't his priority. What matters "is how well the new system will run the science," he said.
The Environmental Molecular Sciences Laboratory keeps what it calls the Green Book, which scientists at the laboratory use to record their expected research needs over the next five to 10 years. In the book, scientists are writing "things like, 'We need five to 10 times more processing power, [and] more memory,'" Regimbal said.
That information served as guidance for purchasing the new supercomputer, he said, adding that the current system is oversubscribed with requests for processing time. That problem likely will arise again with the new system. "In three to four years, the science will have moved on to the point where [the users] will need more than we can provide," Regimbal said.
Also today, IBM is expected to announce that the Oak Ridge National Laboratory in Oak Ridge, Tenn., is buying a second-generation version of its Blue Gene/P supercomputing system with about 2,000 quad-core PowerPC processors and potential peak performance of 27 TFLOPS. The system's price wasn't immediately available.
The Blue Gene/P system was built to fit into two cabinets. Thomas Zacharia, associate laboratory director of computing and computational sciences at Oak Ridge, said that a distinguishing element of the system is its high-speed network, which is helped by the machine's relatively high density.
Oak Ridge has developed some algorithms that require the use of thousands of processors at once. "Even with the speed of light, as you stretch these processors further and further, it takes longer and longer to send messages between them," Zacharia said. He added that packing the processors into a small amount of space reduces latency, enabling Oak Ridge "to solve calculations using the entire 8,000 cores on a single application."
The need for faster and more powerful systems has propelled the high-performance computing market to grow by more than 9 percent annually, with worldwide spending expected to reach US$14.3 billion by 2010, according to figures released late last year by IDC.
Ed Turkel, manager of product and technology marketing for HP's high-performance computing division, said he thinks the supercomputing market is growing at a faster rate than the overall server business is. He also views high-performance computing as an incubator for technologies and new design approaches that may eventually find their way into business systems -- high-speed networking, for instance.