Beyond the need for speed
However, simply owning the fastest supercomputer does not make China a supercomputing superpower by default: its might has to be demonstrated in how well all that power is put to use in applications. Indeed, Tianhe-2 has previously been criticized as a ‘stunt machine’ that has impressive specifications but lacks the software that makes it easy for researchers to use, on top of being too expensive to run.
“The widespread obsession with supercomputer hardware is understandable. No one who helps administer the TOP500 Supercomputer Sites project, as I do, can claim to be immune to it,” Dongarra mused. “But when it comes to advancing the cause of computational modeling and simulation as a new part of the scientific method, there is no doubt that the complex software ecosystem must take centerstage. Software is the potential blind spot for any supercomputer architecture.”
“Focusing on developing a supercomputer that ranks high on the TOP500 list but is only good for solving linear algebra problems is not a good use of funds,” agreed Deng. “China has recognized the value of software development and is investing accordingly.”
As evidence of this investment, three of the six finalists for the prestigious Association for Computing Machinery (ACM) Gordon Bell Prize 2016 were teams that used Sunway TaihuLight. One of the teams, a 12-man group that included Associate Professor Fu Haohuan of Tsinghua University in Beijing, was announced the winner at the 2016 International Conference for High Performance Computing, Networking, Storage and Analysis held in Salt Lake City, Utah in November 2016.
Aside from its raw power, perhaps what makes Sunway TaihuLight so astonishing is that it is not only twice as fast as its predecessor but also three times as efficient. This benefit is derived from relying on modern, China-controlled instruction set architecture instead of the antiquated complex X86, as well as an elegant platform and cooling system.
“On the one hand, TaihuLight uses a fully automated, closed-coupled liquid cooling unit. On the other hand, the core operating frequency for Shenwei SW26010 proccessors is low, at just 1.5GHz,” explained Fu, who is also the deputy director of the National Supercomputing Center at Wuxi where Sunway TaihuLight is housed.
“These factors greatly reduce energy consumption, giving Sunway TaihuLight an efficiency of 6.05 gigaFLOPS/W, which is comparable to other top international systems.”
2020 vision
As impressive as it may be, Sunway TaihuLight is just the tip of the iceberg of China’s supercomputing ambitions.
“High performance computing is recognized as a powerful engine for growth and the key to China’s competitiveness in science and technology,” Fu shared with Supercomputing Asia.
“At the Wuxi National Supercomputing Center, we will make use of the resources provided by Jiangsu province, combined with funding from the 13th Five-Year Plan, to build a globally influential technology innovation center.”
One important area China hopes to take a lead in is building the first exaFLOPS-scale supercomputer. In May 2016, the central government announced that its 13th Five-Year Plan includes funding for a program to develop the country’s first exascale supercomputers—Jinan University’s system, which is based on a similar ShenWei architecture as TaihuLight, as well as the ARM-based Tianhe-3—which are both expected to be ready by 2020.
In the race to exascale, China has serious competition in the form of old rivals Japan and the US. Japan was in fact the first to announce a comprehensive plan to build an exascale computer, launching Flagship 2020 in 2014. Meanwhile, the US is currently working on two computers that could topple Sunway TaihuLight in the near future: IBM’s 200 petaFLOPS Summit supercomputer at the Oak Ridge National Laboratory and the 300 petaFLOPS Sierra at the Lawrence Livermore National Laboratory. In addition, the US Department of Energy has plans for two exascale machines by 2023, according to the Exascale Computing Plan which was made public in April 2016.
“I am quite optimistic that China is on track to reach exascale by 2020 for the following reasons: There is still more than three years to go; China has already completed a 125 petaFLOPS system; it is able and willing to spend a lot of money; and the ranking is based on Linpack, which China is already strong in,” Deng said.
“China is big and supercomputing problems are expensive to solve, but its supercomputing communities are highly collaborative. We will require the cumulative efforts of future generations if we are to develop the necessary central processing units, complete software systems and application packages for these exascale computers to become a reality.”
This article was first published in the print version of Supercomputing Asia, January 2017.
———
Copyright: Asian Scientist Magazine.
Disclaimer: This article does not necessarily reflect the views of AsianScientist or its staff.
