There's a new supercomputing champ in town
... and it's not the Energy Department
- By Henry Kenyon
- Oct 29, 2010
After spending the last few years aggressively pursuing the title, China has now produced the world’s fastest supercomputer. In an announcement made today at the High Performance Computing advisory council workshop in China, the Tianhe-1A supercomputer has set a record performance speed of 2.507 petaflops. It has a peak theoretical performance level of 4.7 petaflops. A petaflop is one thousand trillion mathematical operations per second.
China has developed a number of machines that have steadily pushed operating speeds to match and supersede the world’s fastest supercomputers, which reside in the United States. In June, the Dawning Nebulae computer, housed at China’s National Supercomputing Center, in Shenzen, reached a sustained operating speed of 1.27 petaflops, putting it in the number two place behind the U.S. Department of Energy’s Jaguar supercomputer at the Oak Ridge National Laboratory. Jaguar has a speed of 1.75 petaflops and a theoretical peak capacity of 2.3 petaflops.
China threatens U.S.'s supercomputing supremacy
The Tianhe-1A, which means “River in the Sky” or “Milky Way,” consists of massively parallel arrays of graphics processing units coupled with multi-core central processing units. The system uses 7,168 NVIDIA Tesla M2050 GPUs and 14,336 Intel Xeon processors with an aggregate main memory of 262 terabytes and 2 petabytes of storage in a Lustre clustered file system, according to NVIDIA. Tianhe-1A consists of 112 computer racks, eight storage node cabinets, six communications racks and 14 input/output racks. The Tesla GPUs and Intel processors were also central to the design of the Dawning Nebulae supercomputer.
The extensive use of GPUs also helps reduce the computer’s power demands. According to NVIDIA, a 2.507 petaflop system built entirely with CPUs would consume in excess of 12 megawatts to operate. But the GPUs in the Tianhe-1A allow it to consume only 4.04 megawatts.
Designed by China’s National University of Defense Technology (NUDT), the computer is located at the National Supercomputer Center in Tianjin. It will be operated as an open-access system for running large-scale scientific calculations.
However, while the Tianhe-1A has a great deal of power, what makes it especially important is that it also contains Chinese-designed components. The Register reports that NUDT has developed its own proprietary interconnect for the server nodes called Arch. Future generations of Tianhe machines will also use an indigenous multi-core processor called Godson. According to the Register, the switch in the core of Arch has a bi-directional bandwidth of 160 gigabytes per second, a latency for a node hop of 1.57 microseconds, and an aggregated bandwidth in excess of 61 terabytes/second.
This home-grown capacity also represents an effort to meet national computing demands. The Wall Street Journal reported that the Tinahe-1A is part of an ongoing strategy by China to develop a range of systems to dominate both military and commercial supercomputing applications.