Cray's massive Titan system is now the world's fastest supercomputer, knocking off IBM's Sequoia after only six months in the top position.
The Top500 list of the fastest supercomputers was released Nov. 12 at the SC12 supercomputing show in Salt Lake City, and Titan—a Cray XK7 system installed at the Oak Ridge National Laboratory in Tennessee—took the No. 1 spot with a performance of 17.59 petaflops, or quadrillions of calculations per second. The system boasts 560,640 processors, including 261,632 of Nvidia's Tesla K20x GPU accelerators, which also were announced at the show Nov. 12. It runs Advanced Micro Device's Opteron 6274 processors.
According to Nvidia officials, 90 percent of Titan's performance was delivered by the company's K20X accelerators. Such GPU accelerators from the likes of Nvidia and AMD—which on Nov. 12 introduced its new FirePro S10000 graphics card—are becoming increasingly important in the supercomputer and high-performance computing (HPC) segments. They offer increased computing power and greater energy efficiency when working with traditional CPUs for highly parallel and compute-intensive workloads found in such industries as pharmaceuticals, oil exploration and climate studies.
Intel officials are pushing their upcoming x86-based Xeon Phi coprocessors as alternatives or complements to GPUs in supercomputers, including the massive Stampede system being built at the Texas Advanced Computing Center (TACC) in Austin.
In all, 62 supercomputers on the list leveraged GPU accelerators or coprocessors.
Coming in second on the Top500 list—which is released twice a year—was Sequoia, an IBM BlueGene/Q system at the Lawrence Livermore National Laboratory that hit a performance of 16.32 petaflops. Sequoia's ascension to the top spot in June was an important milestone, bringing back the supercomputing title to the United States after more than three years. Sequoia is powered by more than 1.57 million IBM Power cores.
Claiming the third spot was Fujitsu's K computer at the RIKEN Advanced Institute for Computational Science in Kobe, Japan, which had been the fastest system in the world—with a performance of 10.5 petaflops—until displaced in June by Sequoia. Two other IBM BlueGene/Q supercomputers rounded out the top five—Mira, at Argonne National Laboratory, and JuQueen, at the Forschungszentrum Juelich in Germany. That last system has been upgraded, and is now the fastest in Europe.
The Stampede supercomputers at TACC, running Dell PowerEdge C8220 servers powered by Intel's eight-core Xeon E5-2680 chips and leveraging the Xeon Phi coprocessors, debuted at No. 7 on the list, with a performance of 6.2 petaflops. When it becomes fully operational, Stampede will have a performance level of 10 petaflops, with the coprocessors providing more than 7 petaflops, according to TACC officials.
According to officials that compile the Top500 list, there now are 23 supercomputers that offer more than a petaflops of performance, only four-plus years after the first petaflops system—IBM's Roadrunner—hit the list.
Intel processors are running in 76 percent of the Top500 systems, followed by AMD Opteron chips in 12 percent, or 60 systems. IBM's Power processors are found in 53 supercomputers, about 10.6 percent of the list. Use of the InfiniBand interconnect technology continues to grow, being found in 226 of the supercomputers, up from 209 systems in June. Gigabit Ethernet is used in 188 systems, down from 207 six months ago.
Multicore chips are widely used in the supercomputers on the list, with 84.6 percent using chips with six or more cores, and 46.2 using processors with eight or more cores.
Nationally, the United States is the clear leader in the use of HPC systems, with 251 of the 500 supercomputers found in the country (compared with 252 on the June list). China, with 72 systems, is the second-largest user of supercomputers, followed by Japan, the U.K., France and Germany. Overall, Asian countries have 123 supercomputers on the list—up one from six months ago—while European countries have 105 supercomputers, down one from June.
The performance level to get on the Top500 list has jumped, according to organizers. In June, the entry level for getting on the list was 60.8 teraflops; it's now 76.5 teraflops. To get into the top 100, the performance has jumped from 172.7 teraflops to 241.3 teraflops.