After China’s Shenwei Taihu Lake took the lead in rushing its performance to 100P, Japan couldn’t help it. Japan’s Fujitsu released a PPT to build a 1000P supercomputer, and also cited ARM as a foreign aid. Maybe someone will ask, can it be made?
In fact, Japan's technical strength in supercomputer is not bad, and there have been such a big machine that Beijing once ranked the top of the TOP500. Although it was surpassed by the latecomers on the TOP500, it still tops the list on the Graph500. It’s just that Japan’s economy continues to be sluggish and its fortunes are not good. Japan’s supercomputer is gradually being pulled away by China and the United States.
(Japan)
Japan's super-computing Jing is the world's number one?Just as the light of Shenwei Taihu Lake topped the list at the TOP500, the Nihon Keizai Shimbun reported that "the Shenwei Taihu Lake is ranked relatively high, but its suitable field is limited, which is a difficult machine for users." Japan Many related people said that "supercomputers should be useful."
Professor Matsuoka Satoshi of Tokyo Institute of Technology, Japan -
"The design ideas behind Beijing and Shenwei are very different. Beijing is a supercomputer that can handle a wide range of real-world operations by expanding massive memory. Although the speed of the top 500 is only one-ninth of Shenwei, even Beijing In the face of complex calculations, it is not easy to slow down. Although Shenwei ranks relatively high, it is considered to be suitable for a limited field. Matsuoka said that Shenwei is a more difficult machine for users, which is undeniable."
To this end, the author consulted industry experts, he believes that Matsuoka said that there is some truth, and stated:
"Objectively, if I buy it myself, I choose Beijing."
There are also experts in the industry said -
"After all, heterogeneous many cores are a big trend. . . . . In terms of computing performance, the light of Shenwei Taihu Lake is still the world's number one"
What is unique about Beijing?Although Beijing's performance in Linpack is about one-ninth of that of Shenwei Taihu Lake, although Japan does not currently build a super-computing that rivals the light of Tianhe 2 and Shenwei Taihu, it still has Beijing itself. Its unique.
In July 2016, Beijing won the championship for the third time in the Graph500 rankings.
In the test, Beijing successfully solved the problem of large-scale chart breadth-first search consisting of about 1 trillion root nodes and 16 trillion branch nodes in 0.45 seconds. The benchmark score is 38621GTEPS. Ranked second is China's Shenwei Taihu Lake, with a score of 23756GTEPS; the third is the US Sequoia, 23751GTEPS, and the fourth is the US Mira, 14982GTEPS. What is Graph500? In fact, in addition to the TOP500 that emphasizes Linpack performance, there is also the Graph500, which emphasizes memory bandwidth and latency, and the Green500, which emphasizes energy saving and environmental protection.
Graph500 has high requirements on the system's memory bandwidth and memory latency, and the computing power itself does not affect the test results. This test is biased towards memory-intensive operations, in the context of big data analysis and large-scale chart analysis. More meaningful.
In recent years, there is a view that the Linpack test is outdated. Jack Dongarra, a professor at Tennessee University, one of the founders of TOP500, proposed the HPCG test. Unlike Linpack, which focuses on the computational performance of linear equations, HPCG is more focused on the overall performance balance. The computational and communication characteristics of a class of applications of the conjugate gradient iterative method, but only represent the characteristics of a certain aspect, and can not cover the characteristics of super-calculation. If computing performance and communication are not an order of magnitude, even if the two indicators are strong, they can lead to lower HPCG results. As of now, HPCG has not been widely accepted by the industry.
In the HPCG test, Beijing's performance is also good, second only to Tianhe No. 2, the score is 554.4TFLOPS, Tianhe No. 2 is 580TLFOPS, and the third-ranked Shenwei Taihu Lake HPCG performance is 371.2TFLOPS. If the HPCG/HPL parameter representing the actual efficiency of the calculation is used, the HPCG/HPL parameter of Tianhe 2 is 1.1%, the HPCG/HPL parameter of Shenwei Taihuguang is 0.4%, and the HPCG/HPL parameter of Beijing is 4.9%. . HPCG pays more attention to comprehensive performance, and requires calculation performance, memory bandwidth, delay and other indicators. Once a certain performance is particularly strong, it is a situation of partiality. Even if several indicators are not bad, HPCG operation speed will appear. Low situation.
Imbalances in computing and communication lead to low HPCG/HPLThe reason why the HPCG/HPL of Shenwei Taihu Lake is low is that the imbalance between computing performance and communication is high. The reason why Shenwei Taihuguang can rank first in TOP500, third in Green500, and float in Linpack Point performance 93PFLOPS, Linpack efficiency up to 74.16%, performance and power consumption ratio of 6G / W, to achieve both computing performance and energy saving, and this is still behind the processor manufacturing process Intel two generations, using 28nm manufacturing process The reason for this is that the speed of the operation is increased by a large number of processors in parallel, and the energy consumption is reduced by reducing the memory performance of the large electrode, which will increase the imbalance in the overall performance - this imbalance is not The computing performance is poor, or the Internet is poor. In terms of computing performance, Shenwei Taihu Lake is the top. In terms of the Internet, Shenwei Taihu Lake is only the mainstream level. It is also called the top Internet because of computing and communication. The imbalance caused the HPCG/HPL parameters of Shenwei Taihu Lake to be low.
The reason why Beijing is high in this parameter is because although computing performance and communication are not top-level, because the computing performance and communication are of an order of magnitude, plus the factors that are isomorphic, the HPCG/HPL parameters will Higher.
The significance of these two tests is to illustrate that Beijing's ease of use is better, and by expanding the massive memory, it can widely deal with various real-world operations, has a wide application, and is very balanced in computing performance and communication. In contrast, the light of Shenwei Taihu Lake is a good performance in all aspects, but the type of computing performance is particularly strong, in terms of balance is not as good as Beijing. But this does not mean that the light of Shenwei Taihu Lake is not as good as a big machine a few years ago. In fact, in terms of absolute performance and application in the light of Shenwei Taihu Lake, Beijing can't match it.
What is the next generation of supercomputers in Japan?At the International Supercomputer Conference in Frankfurt, Germany, Fujitsu revealed that its next-generation supercomputer Post-K will use the ARMv8 architecture, which will be launched in 2020 according to the planned Post-K program announced by Fujitsu. The planned application performance is K 100. Times, that is, the performance will reach 1000 PFLOPS, in other words, 10 times the performance of Shenwei Taihu.
In addition, Fujitsu has announced a technical reserve that can be used to build 100P supercomputers - SPARC64 XIfx processor, which has 32 cores for computing and 2 cores for running OS services, somewhat similar to BlueGene/Q. With a 20nm manufacturing process, the dual-precision floating point performance exceeds 1TFlops. As for whether Japan can turn technical drawings into reality, there is only time to test it.
Double Burner Electric Hotplate
Double Iron Hotplate,Double Burner Electric Hotplate,220V Hotplate,Double Burner Hot Plate
Shaoxing Haoda Electrical Appliance Co.,Ltd , https://www.zjhaoda.com