ZT: China taps Nvidia for world's second biggest computer

步從容

Graphics chips boost performance-per-watt for supercomputers
Rick Merritt
EE Times
(05/31/2010 4:00 H EDT)


SAN JOSE, Calif. — Using Intel and Nvidia processors, a Chinese computer maker has installed in Shenzhen what experts say theoretically is the fastest supercomputer in the world. The system represents both the rise of China and of graphics processors as a growing ingredient in high-performance but relatively low power systems.
The Nebulae system built from TC3600 blades computers made by China's Dawning Information Industry Co., Ltd. is the third system to break the petaflops barrier worldwide. It is ranked second on the latest version of the Top 500 Supercomputers list released Monday (May 31) in conjunction with the International Supercomputing conference in Hamburg.
Nebulae was measured with a Linpack performance of 1.271 petaflops/second, below the 1.75 petaflops/s performance of Jaguar at the U.S. Department of Energy’s Oak Ridge Lab that maintains the top spot. However Jaguar, a Cray system based on six-core Istanbul processors from Advanced Micro Devices, has a theoretical peak capability of 2.3 petaflops/s, well below Nebulae's theoretical peak performance of 2.98 petaflops/s.
The system, installed at the National Supercomputing Centre in Shenzhen, uses Intel X5650 processors and Nvidia Tesla C2050 graphics processors as accelerators. The combination of the latest Intel and Nvidia chips are credited with giving the system its muscle.
The graphics processors are also relatively power efficient. The Nebulae system which uses 4,640 Tesla chips, consumes about 2.55 megawatts compared to about 7 MW for Jaguar.
"The performance per power is one of the biggest parts of the graphics story," said Sumit Gupta, manager of Nvidia's Tesla products, speaking of the trend toward general-purpose computing with graphics chips.
Average power consumption of a Top 500 system is 397 Kilowatts and average power efficiency is 195 Mflops/Watt, up from 150 Mflops/Watt a year ago. The most power efficient system on the list also uses graphics chips. The so-called QPace Cluster based on IBM PowerXCell 8i processor blades was rated at 774 Mflops/Watt.
Dawning, the maker of the Nebulae system, has been making high-end computers since 1993 when its Motorola 88000-based system was measured at 640 megaflops/s. Its most recent series 4000 and 5000 systems used AMD Opteron processors.
The company has said it will use in future systems China-designed Loongson 3 processors which have 16 cores each and 8 Mbytes L2 shared cache. However, to date it has not fielded such systems.
A system similar to Nebulae, the Tianhe-1 installed at the National Super Computer Center in Tianjin, pairs AMD graphics processors as accelerators with Intel CPUs to hit number seven on the Top 500 list. Each node of Tianhe-1 consists of two AMD GPUs attached to two Intel Xeon processors.
The performance of Nebulae and Tianhe-1 catapulted China into second place in installed performance at 9.2 percent, still far behind the U.S. at 55.4 percent but ahead of any other European or Asian country. China now has 24 systems in the Top 500, tied at fourth place with Germany whose numbers have been declining steadily