"Milky way, second" super computers each year how much electricity?
Tianhe second-supercomputer-IT news "Milky way, second" super computers each year how much electricity?

Moore thinks IT fields, computer performance will raise 1 time every 18 months. China's "milky way, second" super computer in the world Supercomputer Rankings won the Championship three times in a row, is cause for celebration, but it also shows, a super computer technology into a bottleneck. According to the news website reported on July 3, China computer Federation High Committee Secretary-General, researcher at the Chinese Academy of Sciences calculations said Mr Zhang, super computers by increasing the number of processors in the past will be able to continuously improve performance.

"Milky way, second" super computer

But United States Lawrence · haosi·feimuni Professor at the Berkeley National Laboratory believes that supercomputers not technological developments of recent years, is not enough to break through the barriers to AI ultra-count. Think 2018 will achieve 1000P Ultra is now postponed until 2020 or 2022 will be achieved.

High power processor heat sink

Ultra is currently facing the challenges of development is first and foremost controlling the power consumption. Compared to the current level of integration and has not fundamentally improved, with existing technology, very super computer power consumption of more than 20 MW, the construction cost will be more than $ 200 million, which will bring a series of technical and operational issues. For example, "Tianhe second" only tariffs will be 100 million Yuan a year, operation at full speed, tariffs of up to 150 million. If still under the old way, relying on the increased size and manufacture of extremely ultra-count and power consumption may be up to 50 MW-100 MW, and this requires a dedicated power plant to power it, this ultra is obviously of no practical use. While such a dense array of powerful processors, will also have a cooling problem that cannot be resolved. Internationally recognized standards of extremely ultra-count is that power must be controlled within the 20 MW, this calls for a revolutionary breakthrough in manufacturing process.

Followed by reliability issues, future super computers to be part of more than 1 billion, with the current failure rate on average every 10-20 minutes the system will report a hardware error and handle errors takes half an hour at a time. This is clearly intolerable.

Reports that supercomputers not faced speed restrictions of access to data . The speed of supercomputers now run faster, but existing memory reads data but has been unable to keep up with demand, has also become a super computer performance bottlenecks. United States also has a number of professors ought to reflect on the development of supercomputers not direction, change the mode of pursuit of the peak speed, and focus on practical value. For example give the generic type is, but custom supercomputers not according to actual demand.

Reported that supercomputers not to achieve more long-term development needs to solve architectural problems. State Key Laboratory of high performance server and storage technology Group Director, wave Chief Scientist Wang Endong believed that, and how to develop more innovative system architecture and breakthrough performance bottleneck, performance rose from petaflops to exascale supercomputers not, this is one of the topics of concern to global supercomputer expert. Chinese universities last year in Shanghai Jiao Tong University up and running the first supercomputer using the wave CPU+GPU+MIC, jointly developed with Shanghai Jiao Tong University's innovative technical architecture, the current schema is generally regarded as a breakthrough p levels are possible methods.

Supercomputers not leaps and bounds the United States upset

Throughout the country, China has invested or plans to build eight national Supercomputing Center, which includes a large Supercomputer Center in Wuxi New. Chinese Super quick breakthroughs in the field of, United States uneasy. End of April, the United States Government's Institute on global conflict and cooperation, University of California, San Diego invites China supercomputer experts held a "Supercomputing strategic dialogue between China and" aim is to study Chinese Super now what is the purpose of is used for military and scientific research? United States have been concerned that if Chinese Super core was implemented on hardware independent control, application software also went up, later United States cannot restrict China's progress in the area of supercomputers not, thereby threatening the United States leadership. According to Mr Zhang understood that in fact is now on European and American scientists because they cannot find suitable ultra platform in the country, while the Chinese Super Center to run a research program.