The list is out, and the Chinese have taken the top spot. The new list of the world's top 500 supercomputers, released last week, gives the top spot to the Chinese Tianhe-1A system, which hit 2.57 petaFLOPS, or quadrillion calculations, per second.
The Chinese machine is located at the National Supercomputer Center in Tianjin, and its performance peak was first announced in October. The former number-one supercomputer, the Cray XT5 Jaguar system at the U.S. Department of Energy's Oak Ridge Leadership Computing Facility in Tennessee, is now in second place with 1.75 petaFLOPS/second.
Chinese scientists also hold third place with a contender known as Nebulae. With a peak of 1.27 petaFLOPS/second, it was the former number two. After Nebulae, the next spots are held by the only Japanese machine in the top 10, Tsubame 2.0 at the Tokyo Institute of Technology, with 1.19 petaFLOPS/second, followed by Hopper, another U.S. Department of Energy Cray system. Hopper, an XE6 system, is located at DOE's National Energy Research Scientific Computing Center in California, and it's only the third machine in the U.S. to have broken the petaFLOPS barrier.
Seven of the current top 10 have achieved petaFLOPS status, five are new to the list, and five are in the U.S. The remaining countries represented in the top 10 are France and Germany.
Nvidia graphics-processing units are used in the two Chinese and one Japanese machines, and 17 of the top 500 use GPUs to accelerate computation. Some observers have noted that GPUs are particularly well-designed for the Linpack benchmark used in this ranking, which measures performance in solving a dense system of linear equations. Interestingly, the key components for the top-ranked Tianhe-1A are from two U.S.-based companies, Intel and Nvidia, and it uses a Linux operating system.
In the top 500, the U.S. still holds the most places, with 275, but this is down slightly from 282 in June. China has 42, Japan 26, and India has four. European countries together have 124, and Asian countries 84.
Overall, speed in all the top 10 systems is moving up. The entry level is now 31.1 TeraFLOPS or trillions of calculations per second on the Linpack benchmark, while six months ago it was 24.7. The combined performance of all 500 systems is 44.2 petaFLOPS per second, while just six months ago it was 32.4. In 2009 at this time, the combined total was 27.6.
Data-center managers may be wondering about the power-consumption requirements of these superhero machines. The top 500 list is now tracking this metric, and, although overall consumption is increasing, the efficiency of the systems is also increasing. Only 25 of the top 500 use more than one megawatt, and IBM's prototype of its new BlueGene/Q system has set a new record in power efficiency at 1680 MFLOPS/watt, which is twice the next best system's level.
The top 500/top 10 list is compiled by researchers from the University of Mannheim in Germany, NERSC/Lawrence Berkeley National Laboratory in Berkeley, Calif., and the University of Tennessee in Knoxville. The next list is expected to be released in June.