[China] China, GPUs on the rise in Top 500 list
 
China and graphics processors—especially from Nvidia—are on the rise in the latest list of the world's Top 500 Supercomputers. The latest lists underscores China's continued investment in high performance computing and the spread of hybrid GPU/CPU system designs.

China is second only to the U.S. with 75 of the world's Top 500 supercomputers, up from 64 on the list published six months ago. China also holds bragging rights for systems that rank as the second and fourth most powerful in the world.

The U.S. still leads by far with 263 systems on the list, up from 251. Among other leading countries Japan has 30 systems, the U.K., France and Germany have 27, 23 and 20 respectively.


"China has become a first-level consumer and producer of high performance systems, and I believe they will continue to support it financially while the rest of the world is experiencing tightening budgets," said Erich Strohmaier, an editor of the twice-yearly list.

"My main question is at what level they will stabilize their efforts and if they will be content with being number two behind the U.S.," he said.

China is "producing systems with their own processor designs--they are a force to reckon with," said Jack Dongarra, a computer science professor at the University of Tennessee who also helps compile the list.

As for graphics processors, 39 Top 500 systems now use GPUs as accelerators, up from 17 six months ago. Thirty-five of them use Nvidia chips, two use Cell processors, and two use AMD Radeon GPUs. At least three of the top five systems use a mix of CPUs and GPUs.

The trend in high-end systems comes as even mainstream PC and smartphone processors increasingly combine GPU and CPU cores in a rising tide of so-called heterogeneous architectures. GPUs are credited with handling parallel tasks faster and with less energy than general-purpose CPUs which outperform them in serial computations.

Nvidia is commanding the lion's share of the supercomputer sockets to date because the company "was the first to the market and has a good software base and marketing arm," said Dongarra.

Japan cracks ten petaflops

There were no changes in the ranking of the top ten supercomputers for the first time since the list was first compiled in 1993. But shifts at the top are expected within the next six to 12 months.

Within a year "there will be a number of new entries in the U.S. from Argonne, Lawrence Livermore and Oak Ridge National Labs," said Dongarra.

"There are plans for several 10+ petaflops systems in the U.S., and it will be interesting to see which systems come next and at what levels," said Strohmaier who also is a researcher at Lawrence Berkeley National Laboratory.

While the rankings did not change, the top system on the list, Japan’s K Computer, jumped from 8.16 to 10.51 petaflops on the Linpack benchmark the list uses. The system now uses a whopping 705,024 SPARC64 processor cores from Fujitsu, and does not use any GPUs or other accelerators.

The next closest system, China's Tianhe-1A, hit 2.57 petaflops. The largest U.S. system, called Jaguar and installed at the Oak Ridge National Lab, is third on the list at 1.75 petaflops.

Among other trends the average Top 500 system now uses 18,383 cores, up from 15,520 six months ago and 13,071 a year ago. Currently 62 percent of the systems use processors with six or more cores.

Processor and system vendors had generally stable rankings compared to the previous list.

Intel processors are used in 384 systems (76.8 percent), down from 386 systems (77.2 percent) six months ago. AMD follows with 63 systems (12.6 percent), down from 66.

IBM's Power processors came in third with chips in 49 systems (9.8 percent), up from 45.

IBM kept its lead in systems with 223 computers (44.6 percent) on the list compared to archrival Hewlett-Packard with 140 systems (28.0 percent). Cray, SGI, Bull and Appro follow with 5.4, 3.3, 3.0 and 2.6 percent respectively.

In terms of power use, 29 systems on the list use more than one megawatt. The K Computer that tops the list consumes 12.66 MW, but is one of the most efficient systems on the list at 830 Mflops/watt. The most energy efficient system is the IBM BlueGene/Q at 2,029 Mflops/watt.

Gigabit Ethernet continues to be the most widely used interconnect among the Top 500 which are mainly cluster architectures. However Infiniband is edging toward parity with use in 213 systems, up from 208, compared to 223 for Gbit Ethernet, down from 230.

However, Infiniband-based systems account for almost twice as much performance at 28.7 petaflops compared to Gigabit Ethernet at 14.2 petaflops.

Source: EETimes
 
BACK