Japan’s K Computer passes 8 petaflop barrier, but also sets new power consumption record. The Chinese Tianhe-1A supercomputer, took the number one position last November with a performance of 2.57 petaflops
A new supercomputer from Japan whose performance passed the 8 petaflop milestone ended China’s brief period atop the list of the world’s fastest supercomputers.
The Japanese system also set another all-time Top 500 record — a 10 megawatt power rating while running the Linpack test used to determine system performance. Despite the significant power consumption, the K Computer achieved “extraordinarily high computing efficiency,” said RIKEN and Fujitsu, in a statement.
The new Top 500 leader, the K Computer housed at the Riken Advanced Institute for Computational Science in Kobe, Japan, runs 68,544, eight-core Sparc chips made by Japan-based Fujitsu. The system is expected to eventually run some 80,000 of the Sparc processors.
The Chinese Tianhe-1A supercomputer, took the number one position last November with a performance of 2.57 petaflops.
In speeches after that, President Barack Obama referenced China’s accomplishment.
Supercomputer developers are keenly aware that they will need huge gains in power efficiency to reach supercomputing’s next big goal — building an exascale class system (1,000 time more powerful than a petascale system) by 2018.
“The problem is the trend that power-consumption is increasing,” said Erich Strohmaier, who heads the Future Technology Group of the Computational Research Division at Lawrence Berkeley National Laboratory and is a founder of the Top 500 list.
“Even if it is not desirable, we can adapt to 10 MW for the very largest systems, but we cannot allow power consumption to grow much more,” said Strohmaier. “This has been realized in the U.S. research community for a while and the Exascale initiative of the DOE (U.S. Department of Energy) is addressing this issue directly. Power consumption is already influencing computer design decisions and will have a big influence on the details of future HPC (high-performance computing) architectures.”
The average power consumption of top 10 systems in the latest Top500 list is 4.3 MW, up from 3.2 MW just six months ago.
There is a major focus globally on developing the architecture, software and hardware needed to produce a system capable of reaching supercomputing’s next milestone.
Dave Turek, vice-president of exascale computing at IBM, puts 20 MW as the idea range for power consumption by an exascale system.
Turek’s view was underscored by the DOE. In a report earlier this year, the agency put the “practical power limit” of an exascale system at 20 MW.
Achieving that goal would require a reduction in power consumption per operation in the range of 300 times less than what is now required for petascale systems, the DOE reported.
If an exascale system can be built at 20 MW, said Turek, it would mean that a petaflop can be delivered at 20 KW. Such a system would put the kind of computation power represented by Japan’s K Computer within reach of the typical corporate data center.
Building an exascale system at 20 MW “will be truly transformational to the IT industry, like nothing that has never been done in HPC before,” said Turek.
Turek said the path to creating an exascale system is unlike that which was followed to achieve petascale performance. “We’ve been maybe a little too cavalier as an industry in terms of just extrapolating from the past and thinking that’s a pathway to the future,” he said.
Major innovations will be required in memory architectures, interconnects, optical technologies, and other elements of the system, Turek added. Making the task more difficult is the fact the many of the pieces are built by a wide range of vendors.
The research effort will also require government funding, Turek said.
“From our perspective as a company, we’re going to spend a ton of money in this area — an absolute ton of money to drive [HPC] technology,” said Turek.
IBM has built 213, or nearly 43 per cent, of the systems on the latest Top500 list. It is followed by Hewlett-Packard, which built 153, or about 31 per cent, of the systems on the list.
The increasing global competition by may be helping to change the benchmarks for measuring supercomputing power.
The Linpack test, which measures floating point computing power, has long dominated the market and isn’t going away, but the industry may have to start giving more attention to new benchmarks for measuring the system capabilities. These could include the Graph 500, which ranks data intensive applications, and the Green500, which ranks the world’s most energy efficient systems.
The U.S. has five of the top 10 spots in the latest Top 500 list, while Japan and China each have two and France has one.
U.S. computers also dominate the full Top 500 list with 256, or 51 per cent, of the supercomputers. Next up is China at 62 systems, or about 12 per cent of the total. Germany has the third most systems on the list with 30, followed by the United Kingdom with 27, Japan with 26 and France with 25.