SAN JOSE, Calif. The world officially has a second petaflop-class computer. The Jaguar system at U.S. Oak Ridge National Laboratory hit Linpack performance of 1.059 petaflops using a Cray XT5 system in the latest rankings of the Top 500 Supercomputers published Monday (Nov. 17).
The system ranked second on the list to the Roadrunner computer at Los Alamos National Laboratory built by IBM Corp. that was the first to crack the petaflops barrier in rankings released in June. Roadrunner was slightly upgraded to hit 1.105 petaflops.
In one report, a Cray executive said the Jaguar system was measured as high as 1.64 Pflops. The official rankings were released at the Supercomputing 2008 conference in Austin.
The Roadrunner uses a mix of AMD Opteron CPUs and IBM QS22 blades based on a version of its Cell processor linked on an Infiniband network. The Jaguar uses AMD Barcelona processors a proprietary 3-D Torus interconnect called SeaStar2 developed by Cray.
The current list showed significant stability in the leading chip and system vendors, users and architectures. One standout, however, was the number ten system, the Dawning 5000A at the Shanghai Supercomputer Center based on AMD Barcelona CPUs on an Infiniband network running Microsoft Windows HPC 2008.
Both China and Microsoft have been behind systems that landed in the top 10 before, but neither appears high on the list regularly. The Dawning 4000A, also built by China's Dawning Information Industry Co., Ltd., hit the top 10 spot in June 2004.
"It certainly shows sustained interest, work, and financial support on both parties to establish themselves as players in high performance computing," said Erich Strohmaier, a researcher at Lawrence Berkeley National Laboratory and one of four experts who compiled the list.
The Top 500 list recently started tracking power efficiency of supercomputers. That's because power is an increasingly key issue for large systems often constrained by the amount of electricity they require and heat they can dissipate.
The Roadrunner got top marks in the category, delivering up to 536 Mflops/Watt, thanks to its use of the Cell as an accelerator. The IBM BlueGene/P systems came in second at up to 372 Mflops/W.
Commodity processors are catching up fast. Systems based on Intel's quad-core Harpertown processor ranged from 227 to 265 Mflops/W. Cray systems using AMD CPUs are in a similar range of up to 232 Mflops/W. Average power consumption of a Top 500 system is 358 kWatt, and average power efficiency is 132 Mflops/W.
Intel continues to dominate the Top 500 as a CPU provider in 379 system or just more than 75 percent of the total. IBM and AMD are nearly tied for second place with 60 and 59 systems (12 and 11.8 percent) respectively.
The number of quad-core systems on the list is rising rapidly, lead by Intel Harpertown and Clovertown chips which grew from 252 systems in June to 293 in the current list. A total of 336 systems on the list use quad-core CPUs, 153 use dual-core processors and just four use single-core processors.
Clusters remain the dominant architecture, used in 410 systems. On average systems on the list use 30,490 cores up from 24,400 six month ago.
Gigabit Ethernet is the most used interconnect, appearing in 282 systems. Infiniband ranks second with use in 141 systems. Interestingly no system in the Top 50 uses Gigabit Ethernet, a level at which higher performance Infiniband and proprietary interconnects dominate.
Hewlett-Packard and IBM continue to lead, by far, the list of vendors on the list. HP took the lead in the current ranking with 209 systems (41.8 percent) compared to IBM with 188 systems (37.6 percent). IBM had 210 systems (42.0 percent) six months ago, compared to HP with 183 systems (36.6 percent).
Cray, Dell, and SGI follow with 4.4 percent, 4.2 percent and 3.4 percent respectively.