SAN JOSE, Calif. – A supercomputer in China retained its significant lead as the world's most powerful computer in the latest Top 500 list. Among few changes on the new list, Cray delivered a new system that is now the most powerful in Europe and Infiniband gained in use as a cluster interconnect.
The latest rankings provided more details about Tianhe-2, revealing the sophistication of the supercomputer developed by China’s National University of Defense Technology to be deployed in Guangzho. The system delivers 33.86 petaflop/s on the Linpack benchmark, about twice as much as the world's second most powerful system. It packs 3.12 million computing cores, a mix of two Intel Xeon Ivy Bridge and three Xeon Phi processors on each of 16,000 nodes.
China's engineers developed many of the Tianhe-2 components including its interconnect chips, front-end processors, operating system, and software tools. The system uses 4,096 China-developed Galaxy FT-1500 CPUs, 16-core chips based on the Sparc V9 architecture. Each front-end processor delivers 144 Gflop/s at 65 watts, compared to 211 Gflop/s for the 12-core Intel Ivy Bridge CPUs.
Tianhe-2 runs China-developed Kylin Linux, first approved for use in 2006. It is based on OpenMC, a China-developed programming model, similar to Open-MP, according to details released with the new list at a supercomputer conference.
Cray was the one new entrant to the Top 10 with its Piz Daint at a center in Lugano, Switzerland measured at 6.27 Pflop/s, making it Europe's most powerful system. The Cray XC30 uses 73,808 Nvidia K20x accelerator cores and is the most energy-efficient system in the Top 10, consuming 2.33 MW and delivering 2.7 Gflops/W.
The list now contains 31 petaflop-class systems, up from 26 six months ago. A total of 53 systems use co-processors, 38 of them Nvidia graphics chips, 13 Intel Xeon Phi processors and two AMD Radeon GPUs.
Intel extended its dominance as CPU supplier slightly in the latest list powering 412 (82.4 percent) of the systems, up from 80 percent six months ago. AMD's Opteron chips are used in 43 systems (9 percent), down a point, and IBM Power processors are in 40 systems.
IBM's BlueGene/Q continues to have bragging rights as the most popular architecture in the top 10, used in the No. 3, 5, 8, and 9 systems. However, Hewlett-Packard supplied more systems, 195 percent or 39 percent of the Top 500, compared to 166 systems (33 percent) for IBM.
Clusters continue to scale out. The average concurrency system packs 41,434 cores, up from 38,700 six months ago and 29,796 one year ago.
InfiniBand technology is now found on 207 systems, up from 203 systems. Ethernet, long the dominant cluster interconnect, remained flat at use in 212 systems down from 216 systems a year ago. A total of 77 of the systems now use 10Gbit Ethernet, the rest using Gbit links.
The Top 500 list is compiled by Hans Meuer of the University of Mannheim, Germany; Erich Strohmaier and Horst Simon of Lawrence Berkeley National Laboratory; and Jack Dongarra of the University of Tennessee, Knoxville. The first list was compiled in June 1993.