"Futadake" achieves 4 crowns of the supercon world performance ranking

　The supercomputer at RIKEN has a big difference in the second place in all categories: "TOP500", "HPCG (High Performance Conjugate Gradient)", "HPL-AI", and "GRAPH500" in the supercon world ranking.I got the leader.Will be the leader in June 2020 and November 2020.This is the first time since the share started on March 9, 2021.

　It is currently announced at the international conference on the HPC (high performance computing), which is currently being held online, and will be officially announced on July 1 for the GRAPH500 leader.

　This time, the full specifications of 432 housing and 158,976 nodes are used, and in the RIKEN Institute, "the first place in the third consecutive year of all four rankings is the overall performance of the overall performance.In Society 5.0, which indicates the height of a super -smart society that produces new value, to accelerate the solution of social issues by simulation, and accelerate technological development on AI development and information distribution and processing.As an information -based technology, it demonstrates that it can be fully accommodated. "

　This is the 57th announcement of the TOP500, a performance ranking of supercomputers around the world.As in November 2020, full -spec measurements have 442.01 PFLOPS (Peta Flops), execution efficiency of 82.3%.

　The second place is 148.6pflops in the US "SUMMIT".Has 2.97 times the performance difference in second place.

　LINPACK is a compilation -based solution program developed by Dr. Jack Dongara, a US University of Tennessee in the United States, and the TOP500 is an indicator of execution performance when processing the program.

　The computer's calculation speed ranking is determined by measuring the double -precision floating -point number of arithmetic capacity used in many science and technology calculations and industrial applications.It is positioned as an index that comprehensively shows calculation capabilities and reliability.

　In the "HPCG" ranking of the processing speed of the co -slope method often used in actual applications, including industrial use, the performance of 16.00 pflops is demonstrated.The performance difference was about 5.46 times compared to 2.93 Pflops of the second place Summit.

　HPCG was developed by Dr. Dongara and others who developed LINPACK, and was officially announced as a formal ranking by the international conference "SC14" on the HPC in November 2014 after the announcement of the results of the measurement at ISC14 in June 2014. ing.

　It is one of the methods to converge to the correct solution by performing repeated calculation, and a benchmark program that measures the processing capacity by combining pretreatments to converge to the correct solution quickly.The performance of the computer simulation can be shown.

　"HPL-AI", a performance benchmark for monetary degree and semi-precision operational processing used in deep learning of AI, recorded 2.004eflops.The difference between the second -place Summit has improved its performance compared to November 2020, and the difference from 1.15EFLOPS was about 1.7 times larger.

　In recent years, HPL-AI has improved LinPack benchmarks, mainly Dongara, with a supercomputer equipped with a computer that performs low-precision operations with a dedicated chip for GPU and AI.It has been used since November 2019.It is a benchmark program that has been recognized to solve it in a low -precision operation, and consists of a two -stage calculation process.

　The "Graph500", a performance ranking for large -scale graph analysis, achieved 1002,955 GTEPS.Compared to 23,756 GTEPS of SUNWAY TAIHULIGHT in China, it was 4.33 times the performance difference.

　The Graph500 was worked on a joint research group by RIKEN, Kyushu University, Fix Stars, and Fujitsu, and from 35 trillion branches, about 2.2 trillion peaks and 35 trillion branches due to optimization of communication performance.It is said that the structure -praised first search problem for the composed ultra -large graph was solved in harmony average 0.34 seconds.

　The performance of large -scale graph analysis is an important indicator in the analysis of big data that requires large -scale and complex data processing, but the "Graph500" handles large -scale graphs that exceed 1 trillion., It is necessary to distribute the graph data in multiple nodes.For systems with large -scale networks, optimization of communication performance is also an important factor.

　The joint research group is developing software that can analyze large -scale graphs on supercomputers, and so far, "efficiently split graph data between multiple nodes", "Redundant graph".Developed advanced software technology such as algorithms that reduce search and "optimization of communication performance in large -scale networks of supercomputers".It is said that the use of these has achieved the world's highest performance.

　The joint research group plans to work on reducing redundant search and uniformation of memory use in each node in addition to further optimization of communication performance.

　According to Satoshi Matsuoka, Director of the Research Center for RIKEN, said, "This is a research and development of Japan's best IT technology in Japan, which balances performance, power saving and the ease of programming.be.

　Before the start of common use, not only is the best in the world with simulation, big data, and the main benchmarks of AI for the second consecutive term, but also contributing to the establishment of the government and corporate infection guidelines for new colonovirus measures., In the field of infectious diseases, it has achieved DX.

　This time, it was the third four crowns, indicating a global advancedness in a wide area.RIKEN will continue to sophisticated and develop so that the power can be contributed to Society 5.0 and DGS. "

　"I think that the reason why RIKEN and Fujitsu's efforts to improve application performance were highly balanced is that the reason why the four crowns have been won in the third consecutive term have led to a highly balanced performance that has been highly balanced. ing.

　Fujitsu also plans to utilize it in joint research on drug discovery with the University of Tokyo's Advanced Science and Technology Research Center.We hope that the results obtained by utilizing the use of drug discovery, safe, secure society, and the realization of Society 5.0. "