Global Sources
EE Times-Asia
Stay in touch with EE Times Asia
EE Times-Asia > Embedded

China leads supercomputer top 500

Posted: 18 Nov 2010 ?? ?Print Version ?Bookmark and Share

Keywords:supercomputers? ranking? processors? GPUs? CPUs?

China has kept its number one spot in the 2010 Top 500 supercomputers list, and it extends its reach deeper on the list with 42 of the world's most powerful computers. It is now second only to US in number of supercomputers, with more supercomputers than France, Germany, Japan and the UK.

The rankings show that while Intel's x86 processors still dominate the list, archrival AMD has made some gains. Also, more supercomputers--28 on the latest list--pair such chips with GPUs to boost performance and keep a lid on power consumption.

These hybrid systems include Tianjin's Tianhe-1a, the world's most powerful system measured at 2.5 petaflops on the Linpack benchmark. The third highest performance system at 1.27 Pflops is also a hybrid, the Nebulae built by Dawning Information Industry Co. in Shenzhen.

"I think there will be a lot of discussion of the Tainhe-1a because it�s the first system at the top both to use GPUs and to come from China," said William Dally, chief scientist at GPU maker Nvidia whose chips power the system. "The machine is an impressive development in absolute performance and in the homegrown technology in its network and physical design," said Dally.

Researchers behind the Tianhe-1a have yet to reveal details of Galaxy, the proprietary 160 Gbps interconnect that helps fuel its performance. "I am very interested in how they attacked the more nuanced issues," said Dally who for nearly twenty years has helped design interconnects for U.S. supercomputer company Cray.

The Tianhe-1a "is an engineering feat--we know how hard it is to build a petascale system, but I can't comment on how sustainable its performance is" under real workloads, said Barry Bolding, VP of the products group at Cray.

Researchers have yet to reveal bandwidth and latency figures for Galaxy when the interconnect is under stress of running scientific applications. Details such as random ring latency in such conditions are more telling than the maximum theorthetical throughput China reported, said Bolding.

The Linpack benchmark used as the basis for the Top 500 list itself has been widely criticized as limited even by the researchers who maintain the list.

"People have used Linpack because it is convenient and relatively easy to run, but Linpack is not very representative of machines today," said Dally. "It�s a good proxy for dense linear algebra, but some apps don�t use that and supercomputer workloads are becoming more diverse," he added.

The U.S. remains the leading supercomputer user with 275 of the top systems, down from 282 on the June list. Europe's share dropped to 124 systems from 144. Germany and Japan tied for the third spot behind China with 26 systems each.

Hybrid systems increase
Hybrid systems that mix traditional processors and graphics accelerators are on the rise because they have a performance-per-Watt advantage over systems using just CPUs. However, the hybrid systems are also more difficult to program.

Three of the current top five are hybrid systems, sporting as much as twice the flops/W ratio as the other top five systems which use only CPUs, said Dally. The efficiency of hybrids is important because supercomputers are on track to exceed any reasonable power consumption levels before they hit the next big milestone of exascale performance.

"Ultimately, [hybrid systems] will wind up dominating the list because they are much more energy efficient," Dally said.

A supercomputer under development at Virginia Tech may exceed a Gflop/Watt because it uses a higher ratio of graphics to CPU chips, he said. IBM�s BlueGene/Q system set a record in power efficiency on the latest list at 1.68Gflops/watt, more than twice that of the next best system.

Average power consumption of the latest top ten systems is 3.2MW, up from 2.89MW six month ago. Their average power efficiency is 268Mflops/W, an improvement from 300 Mflops/W six month ago.

The rising power problem is one reason why companies such as Cray have said they will adopt GPUs in future systems such as its XE6 and Cascade supercomputers.

Cray has said it will adopt Nvidia's Fermi chips and its CUDA software to program them. But ultimately it is pushing for development of OpenMP software that can be used with any graphics chips.

"CUDA is very powerful and useful in short term, but we think OpenMP will be more useful in the long run," said Bolding. When it is complete, a hybrid version of OpenMP "can be used with any graphics processor, and Cray would like to stay agnostic so we can pick and choose the best from any generation," he said.

Bolding said he does not think hybrid systems will become widespread until such software becomes available. Meanwhile, Cray also will use in its hybrids compiler technology repurposed from the days when it developed its own vector processors which were in some ways similar to today's merchant graphics chips.

Of the 28 hybrid systems 16 using the IBM Cell processor, ten use Nvidia chips and two use AMD's Radeon GPUs. Dally said use of Cell will decline over time because IBM announced it will not continue the line.

AMD, Infiniband get more use
Among traditional processors AMD made gains over Intel which still dominates the list. Cray edged up in one measure over Hewlett-Packard as a system supplier, and Infiniband made advances on Gigabit Ethernet which is still the leading interconnect.

Intel chips were used in 398 systems, down from 408 systems six months ago. AMD's Opteron family is used in 57 systems, up from 47, while IBM's Power declined from use in 42 down to 40 systems.

Quad-core processors were most prevalent, used in 364 of the systems, 90 percent of them using Intel parts. Processors with six or more cores appeared in 95 systems while dual core chips were used in 37 supers.

AMD's gains came in part from the fact it leapfrogged Intel in the last seven months to deliver some of the first x86 server processors to use eight and 12 cores. Supercomputer designers are quick to grab for their systems processors that sport the fastest data rates or most cores.

IBM kept its lead as a systems vendor with 199 supercomputers, nearly 40 percent of those on the list. HP held second place but slipped from 198 to 185 systems (37 percent) on the list. Cray, SGI, and Dell followed with 5.8 percent, 4.4 percent and 4.0 percent respectively.

IBM also remained the leader with 27.3 percent of total installed performance. Cray jumped over HP to gain the second spot with 19.1 percent of aggregate performance, up from 14.8 percent on the June list.

Almost all the systems on the list are now clusters that use fast interconnects to link increasingly large numbers of multicore processors. The average system on the current list uses 13,071 cores per system, up from 10,267 six months ago and 9,174 one year ago.

Gigabit Ethernet is still the most widely used interconnect, appearing on 227 systems, down from 244 systems. Infiniband appears on 214 systems, up from 205 systems six months ago, mainly focused on higher-end systems.

The next big shift in interconnects is expected when 10BbE interconnects become popular. In the meantime, some observers expect to see a rise in proprietary technologies such as China's Galaxy.

- Rick Merritt
EE Times

Article Comments - China leads supercomputer top 500
*? You can enter [0] more charecters.
*Verify code:


Visit Asia Webinars to learn about the latest in technology and get practical design tips.

Back to Top