The SC09 conference kicks off in Portland, Oregon today with the release of the Top 500 list of supercomputers. It’s a list of computers you’ll never need in your house, or likely even your company, but is the source of significant bragging rights none the less.
The list is compiled by a group of supercomputing experts in the U.S. and Europe and updated twice per year. Aside from the fact it changes constantly, the performance bar shows what Moore’s law has done for computing.
AMD (NYSE: AMD) is proud to say it is in the top machine again, just as it was on the last two lists. Topping the list is the former number two machine, a big kitty called Jaguar at the Oak Ridge National Laboratory. It’s a Cray system with more than 10,000 Opteron processors.
Roadrunner, the former number one, is now number two. It runs a combination of Opterons and IBM Cell B/E processors, which are used in the PlayStation 3. At number three is Kraken, an 18,000 processor Cray XT4 running Opterons at the National Institute for Computational Sciences (NICS) on the campus of the University of Tennessee.
All told, five of the top ten and 10 of the top 20 machines are AMD-powered, noted John Fruehe, director of Opteron product marketing at AMD. “So people across the board are coming to depend on AMD for supercomputing needs,” he told a conference call of reporters.
Of course, Intel (NASDAQ: INTC) is the king of the hill overall, with more than 400 of the top 500 machines, and will likely reign supreme again after six more months of Nehalem deployments. AMD was well-represented in the top of the list but only had 42 machines on the 500 list this time.
All of these AMD systems were upgradeable, since AMD has released several generations of socket-compatible processors. Roadrunner, for instance, was initially deployed with the first generation of Barcelona-era quad-core Opterons. The Department of Energy has since upgraded those machines to newer, six-core Istanbul Opteron processors.
Next generation Opterons
AMD showed a video (here on YouTube) at its analysts day meeting earlier this week, showing the upgrade process can be done in minutes.
Fruehe said that beginning with the forthcoming Magny-Cours 12-core platform, AMD goes to a new socket that will be compatible with its next-generation core, Bulldozer, and promises a few years of upgrades as well.
“So the upgradability that AMD delivers in getting the Jaguar and Kraken system in number one and number three slots will be repeated,” he said.
The original Opteron platform dates back to 2006 and is getting long in the tooth. The next generation will have four DDR3 memory channels instead of two DDR2 channels, so memory performance will increase significantly, he said. Also, the new platform will support PCI Express v2, which means faster I/O as well.
Supercomputing customers currently bidding on a Magny-Cours platform are also including upgrades to Interlagos, the next generation processor, as an upgrade, so customers are actually buying a few generation’s worth when they buy a system.
GPUs go to work
They are also looking at GPUs. This list, the first include systems powered by graphics processor, includes a Chinese supercomputer, Milky Way, at the National University of Defense Technology (NUDT). It cost 600 million Yuan, which translates to US$87.88 million, a real bargain by U.S. terms.
Chinese scientists created a supercomputer with 6144 Intel Xeon E5540s and 3072 Xeon E5450s processors and 2560 ATI Radeon HD 4870 X2 cards. Together, the system can reach a theoretical speed of 1.206 petaflops, which would exceed the best of the most recent Top 500 list and make it the fastest in the world.
Even the Chinese were stunned at their creation. “I was shocked at the milestone breakthrough, which was beyond expectation,” Zhang Yunquan, a researcher with the Institute of Software of the Chinese Academy of Sciences (CAS), told China’s People’s Daily publication. “I previously forecast China’s first petaflop computer no earlier than the end of 2010.”
Fruehe said he wasn’t thrilled to see Xeon’s in the machine, but figures there will be more GPU-powered supercomputers. “From my conversations in the past six to twelve months I have rarely found a customer who is not interested in GPU and is not looking at how do they integrate that in,” he said.