Friday, March 11, 2011

China's Great Leap Forward in Supercomputing

China's Great Leap Forward in supercomputing has not been a secret, especially after the Top500 list published November 2010. Made-in-China supercomputers dominated top 3 with No. 1 Tianhe-1A in Tianjin and No. 3 Nebulae in Shenzhen. Tianhe-1A's 2.566 PFLOPS sustained power dwarfed the output of 1.759 PFLOPS of the No. 2 Jaguar from the US. What is more impressive is the energy efficiency. Tianha-1A achieved the mighty power with 4040 KW electricity while the Jaguar which is based on traditional architecture required 6950 KW to lingering behind. Chinese systems make use of a GPGPU accelerated computing method pioneered by NVIDIA.

The No. 4 on the chart, Tsubame from Japan is also interesting. Its 1.192 PFLOPS output is about 2/3 of the Jaguar, but it consumes only 1/5 of the energy.

What is worth writing, however, is the slow and hesitated response on the US side. All four largest supercomputing systems in the 10 PFLOPS scale will be based on legend technology, which rely on massing tremendous historical CPUs to work collectively. Perhaps a partial explanation is that the US is far advanced in CPU design and manufacturing business, while China is terribly lacking behind not only the US, but also pretty much everyone else. The problem of the US's strategy is that you can only beat the old horse to an extend. To make an analogy, you can improve a propeller plane so amazingly, but it still won't beat a primitive jet plane.

Although it's still too early to call the game, but China's risk taking approach already make an impact in China's computing industry and research. For the first time, the majority top papers (those got referenced most by other researchers, thus deemed most valuable contributions) in a broad cutting edge research domain are produced in China. Mainland based researchers and programmers had never been so confident in their skills and capabilities, which will have lasting effect. The initial success has been tripling down to related areas too. China just announced new supercomputers with Made-in-China CPUs, which will mark the first ever time a Chinese CPU would be put in a serious use.

Japan launched a similar effort in the 1980s when the Japanese government organized a project to study the next generation (5th generation) computers which weighs more on logic deduction than on computation. The project did not reach its promised goals. However, it left us a wealth of knowledge in the areas of logic computing and parallel computing.

China's GPU supercomputers are different than Japan's Fifth Generation Computers. The latter was like setting up a goal of interstellar traveling, then organize scientists and industry power to make it happen. The current Chinese tide is more likely quietly setting up a base on the Mars, then digging around to see what might be valuable. In other words, they can't loose, even if they do not win. The question is, when will the US make up its mind to play the catchup game.

No comments: