China, Nvidia Build World's Fastest Supercomputer
Despite all of Nvidia's talk in recent years about the potential for GPUs in supercomputing, when the Top 500 supercomputer list came out last November, it was rival AMD that made the list with a Chinese supercomputer called Milky Way, which used a combination of Intel Xeons along with 2560 ATI Radeon HD 4870 X2 cards to build the first petaflop computer outside the U.S. and the first to use graphics processors. After years of talking up GPUs in high performance computing, Nvidia has some petaflops to finally back it up.
Six months later and Nvidia (NASDAQ: NVDA) is on top, thanks to another Chinese monster computer called Nebulae. The blade system uses Intel X5650 processors and Nvidia Tesla C2050 GPUs to achieve a theoretical peak performance of 2.98 petaflops per second, or PFlop/s, according to the latest Top 500 list of supercomputers in the world.
That makes it the fastest supercomputer in the world when measured by peak performance. But in terms of sustained performance as measured by Linpack benchmarking, it comes in second, with 1.271 PFlop/s of sustained performance.
The top dog, or in this case top cat, remains Jaguar, a Department of Energy supercomputer at the Oak Ridge Leadership Computing Facility. Its top sustained speed is 1.75 petaflop/s, the same as it was six months ago when the November 2009 list came out. Its theoretical peak is 2.3 petaflop/s.
The TOP 500 list is issued twice yearly, in June and November. It was formally presented at the International Supercomputer Conference '10 taking place in Hamburg, Germany.
Nvidia was subdued in its victory. "Tesla has fast become the must-have tool for scientists all over the world. The amount of progress being made in the areas of science, energy, finance and supercomputing is astounding," said an Nvidia spokesman in an e-mail to InternetNews.com.
Roadrunner, the Los Alamos system that ran a hybrid of AMD (NYSE: AMD) Opteron processors and IBM (NYSE: IBM) Cell processors and was the first ever petaflop system in June 2008, dropped to No. 3 with its performance of 1.04 petaflop/s, unchanged from the last list.
At No. 4 is Kraken, a Cray XT5 system at the National Institute for Computational Sciences (NICS), with performance of 832 teraflops.
Kraken is used in a variety of science programs for things like climate and weather modeling and improving efficiency of biofuels. In fifth is the Jugene supercomputer in Germany, built on IBM's BlueGene/p supercomputer technology and No. 6 is Pleiades, a supercomputer at the NASA's Ames Research Center built by SGI, the highest-positioning computer by the firm. Milky Way, called Tianhe-1 in China, fell to seventh on the list.
The list is still owned by Intel (NASDAQ: INTC). Its x86 architecture runs 401 of the highest-performing systems, and another five are powered by Itanium. AMD has 49 systems, IBM holds 42 with its Power RISC processor, and Oracle's Sparc has two systems on the list. The Xeon 5500 line is the fastest growing, with 186 systems. It was in just 95 on the November 2009 list.
Quad-core processors are used in 85 percent of the systems and another 5 percent use processors with six or more cores. Intel began shipping eight-core Xeon 7500 processors in March and AMD shipped its eight- and 12-core processors around the same time, so they could show up as soon as the November 2010 list.
IBM has 196 systems while HP (NYSE: HPQ) is right behind it with 186. Cray had 21, Dell and SGI were tied with 17 and Oracle/Sun had 12. Cray's systems are very popular with researchers, and 10 of its 21 were in the Top 50 of the list.
Generic Linux also dominates the list, powering 405 of the systems. Various permutations of Red Hat and SuSE Linuxes accounted for another 44 systems, followed by IBM's AIX Unix OS with 19. Windows HPC has just five systems.
The ISC '10 show runs until June 3.