K computer

from Wikipedia, the free encyclopedia
One of the cabinet prototypes of the K Computer

The K computer ( Japanese 京 , Kei , [ keː ], German "10 quadrillion, 10 peta -") is a Japanese supercomputer . With a computing power of 10.51 peta flops (measured according to the LINPACK benchmark) it was considered the fastest computer system in the world in November 2011.

The theoretical maximum power is 10.51 petaflops. It was built by Fujitsu and is located in Kobe at the RIKEN Advanced Institute for Computational Science . It combines 88,128 SPARC64 VIIIfx CPUs (2.0 GHz) with 8 processor cores each in 672 cabinets. Linux is used as the operating system . The expansion of the system should be completed in November 2012 with 864 cabinets and an output of 11.28 petaflops.

As early as June 2011, the K computer in its first expansion stage (8.16 PFLOPS) replaced the Chinese Tianhe-1A (2.56 PFLOPS) as the world's most powerful supercomputer up to then. In June 2012, this record was broken by the US IBM Sequoia (16.3 PFLOPS) , which is about twice as fast .

The successor to the K computer is the Fugaku , which was set up in May 2020.

6-dimensional mesh / torus topology network technology (Tofu Interconnect)

The K-Computer is a large system with more than 80,000 CPUs. The network that carries data such as B. Exchanging calculation results between the CPUs plays a very important role. The K-Computer Network, called Tofu, uses an innovative structure called the "6-Dimensional Mesh / Torus" topology. This offers many communication paths between the neighboring CPUs. By executing the data communication between the CPUs on the shortest route and in the shortest possible time, it is ensured that the network enables the CPUs to have full computing power.

In order to always achieve the highest performance, it is still important to prevent failures. Even if there is a partial failure of components, the impact must be minimized. This is achieved by configuring alternative routes in the network between the CPUs. A mechanism that bypasses faulty CPUs ensures that data exchange can continue. The computational processing is not interrupted.

Simulation of the human brain

On August 2, 2013, the Japanese research institute Riken announced that it had carried out the largest simulation of the nervous system of the human brain to date in cooperation with the German research center Jülich . The virtual neural network to be simulated consisted of around 1.73 billion nerve cells, which were connected to one another by around 10.4 billion synapses. The 82,944 CPU strong computing cluster K required 40 minutes of computing time to simulate only about 1% of the brain activity for 1 second. In order to obtain an exact mathematical model, a 24-byte memory was allocated to each synapse, which led to a 1-petabyte working memory consumption. Team leader Markus Diesmann was very confident about the development of simulating the brain as a whole at neuron level: “If Peta-scale computers like K can represent 1% of the network structure today, then [...] will be with Exa-Scale Computers can safely simulate the whole brain within the next decade. ”The open source project NEST was used as simulation software.

See also

  • TOP500 list of the 500 fastest computer systems (updates every six months)

Web links

Individual evidence

  1. ↑ 次世代スーパーコンピュータの愛称は「京(けい)」と決定-10ぺタを表し,京速(けいそく)コンピュータとしてなじみがあることなどが選考の理由- . RIKEN, July 5, 2010, accessed June 22, 2011 (Japanese).
  2. ^ "K computer" Achieves Goal of 10 Petaflops
  3. ↑ K computer, SPARC64 VIIIfx 2.0GHz, Tofu interconnect. In: TOP500.org. Retrieved August 5, 2016 .
  4. ↑ 資料 5 成果 報告 票 (そ の 2) . (PDF; 1.6MB) In: Ministry of Education, Culture, Sports, Science and Technology of Japan. Retrieved August 5, 2016 (Japanese).
  5. ↑ Fujitsu press release
  6. ↑ http://www.top500.org/lists/2011/06/
  7. ↑ http://www.top500.org/lists/2012/06/
  8. ↑ Innovative "6-Dimensional Mesh / Torus" Topology Network Technology
  9. ↑ Interconnect of K Computer (PDF; 2.4 MB)
  10. ↑ Tofu: A 6D Mesh / Torus Interconnect (PDF; 2.3 MB)
  11. ↑ Tofu: A 6D Mesh / Torus Interconnect for Exascale Computers ( Memento from December 1, 2009 in the Internet Archive )
  12. ^ An Interconnect Controller for the Tofu Architecture. (PDF; 2.6MB) In: Fujitsu. August 24, 2010, accessed August 5, 2016 .
  13. ↑ Current Status of FEFS (PDF; 2.2 MB)
  14. ↑ Programming on K Computer (PDF; 395 kB)
  15. ↑ HotChips Video 22 "Interconnects" on YouTube
  16. ↑ Largest neuronal network simulation achieved using K computer