Ranking of supercomputers in the world Top500

The main articles are:

• Supercomputers
• Rating of supercomputers of Russia and the CIS Top50

2023: There has been a change in the top three most powerful supercomputers in the world

On November 13, 2023, the 62nd edition of the Tor500 rating was published - a list of the 500 most productive computing complexes in the world. The most powerful supercomputer on a global scale, as a year earlier, is Frontier, installed at the US Department of Energy's Oakridge National Laboratory (ORNL). The performance of this system reaches 1.194 Eflops.

At the same time, there have been changes in the first three. So, in second place in the Tor500 rating is a newcomer - the Aurora supercomputer, which is located in the Argonne National Laboratory (ANL) of the US Department of Energy: the performance of this complex is 585.34 Pflops. In the future, the history of Aurora is planned to be brought to 2 Eflops. Another new representative of the list closes the top three - the Microsoft Azure Eagle cloud supercomputer, which showed a result of 561.2 Pflops. The Fugaku (Institute for Physical and Chemical Research of Japan) and LUMI (Euro HPC/CSC, Finland) systems, which were previously on the second and third lines, dropped to fourth and fifth positions with 442.01 Pflops and 379.7 Pflops, respectively.

The United States and China received most of the lines in the Tor500 rating. Thus, the United States accounts for 161 computing complexes, while the PRC operates 104 of the world's most powerful supercomputers. In North America as a whole, there are 171 complexes from the list, in Asia - 169, in Europe - 143.

It is noted that the most popular processors in LDC systems are chips from Intel, AMD and IBM. So, in the top ten, five systems use Intel Xeon products (Aurora, Eagle, Leonardo, MareNostrum 5 ACC and EOS Nvidia DGX SuperPod), and two more complexes are based on AMD solutions (Frontier and LUMI) and IBM (Summit and Sierra).^[1]

2021: Yandex and Sberbank brought Russia into the top ten countries in terms of the number of powerful supercomputers

On November 15, 2021, a list of the top 500 most powerful supercomputers in the world was published, updated twice a year. This time it included 7 Russian supercomputers, while in the previous, June, only 3 systems were included in the list from Russia. The country's positions on the list have improved thanks to the supercomputers of Yandex, which had not previously submitted its systems to the rating, and Sberbank.

Yandex's supercomputer Chervonenkis took 19th place in the top 500 supercomputer rating, becoming the most productive system not only in Russia, but throughout Eastern Europe. Its actual performance is 21.53 Pflops (quadrillion floating point operations per second).

Yandex supercomputer Chervonenkis

In addition to Chervonenkis, the top 500 includes two more Yandex supercomputers - Galushkin and Lyapunov. They ranked 36th in and 40th respectively. The real performance of these machines declared in the list is 16.02 and 12.81 Pflops.

Yandex's new supercomputers are named after Soviet and Russian scientists who contributed to machine learning theory and computer science. According to the company, Lyapunov was commissioned in December 2020, and Chervonenkis and Galushkin in June 2021. The systems are based on AMD EPYC processors and Nvidia A100 graphics accelerators and are used to train neural network models with billions of parameters.

To take measurements, you need to stop the main work with clusters for a day - in fact, model training. We could not afford this even for a day, so we waited for the supercomputers to be enough to take measurements and not stop the work of employees, "Yandex answered TAdviser's question about why they did not submit their systems to the top 500 rating earlier.

A new Sberbank supercomputer has also appeared in the top 500, the launch of which the company announced shortly before in November. With a real power of 11.95, "Cristofari neo" Pflops was ranked 43rd on the list. Its performance is 1.8 times higher than the first Christofari supercomputer, introduced two years ago. He in the ranking is located on the 72nd line.

The Lomonosov-2 supercomputers, located at Moscow State University, and MTS Grom, owned by MTS, occupy 241 and 294 places in the ranking with a peak performance of 2.478 Pflops and 2.258 Pflops, respectively.

The last time 7 Russian systems were present in the top 500 ranking in June 2016, and recently there were only 2-3 domestic supercomputers on the list.

Even after improving its position in the world ranking, Russia is far behind the countries leading the list - China and the United States - both in the total number of supercomputers and in the performance of the most powerful of them. China is on the list with 173 systems, the United States with 149 supercomputers.

At the same time, Russia became the 8th country in the list of the total real performance of the supercomputers presented in it and the 9th in the total number of supercomputers. Russia has the same number of systems in the ranking as South Korea.

The leader of the list remains the updated Japanese supercomputer Fugaku, 7.6 million cores of which are issued by 442 Pflops. And it is still three times faster than its closest competitor Summit.

2020:2 systems from Russia remain in the top 500 supercomputers

supercomputers Two Russian systems remained in the Top500 rating published in June 2020 - created, "Christofari" Sberbank as well as the "company Lomonosov-2 production" T-Platforms established in. MSU

The editorial office of the rating in November 2019 included three systems from the Russian Federation: in addition to the above supercomputers, it was also on the list. Cray XC40 Roshydromet

The most powerful supercomputers by June 2020

The most powerful supercomputer in June 2020 was the Japanese Fugaku system. It has a capacity of 415.5 petaflops, 2.8 times higher than the previous leader Summit, who is now in second place. The new computer also has a high, though not record (ninth on the list), energy efficiency: it is 14.665 gigaflops per watt of power consumed.

The basis of the Japanese supercomputer was the 48-core single-chip ARM systems Fujitsu A64FX, and the total number of cores in the supercomputer was almost 7.3 million.

Experts called this Fugaku supercomputer the best in three more categories:

• when used in industry;
• when used in the field of artificial intelligence;
• when analyzing large digits.

It is noteworthy that before that, not a single computer in the world took first place in four out of six nominations at once. The supercomputer from Japan became the first in the ranking for the first time in nine years after its predecessor, the K computer.

At the same time, Japan remains a small player in the supercomputer market. For comparison, China has 226 systems on the Top500 list, and 114 in the United States.

Chinese supercomputers occupy 226 positions in the list. In second place are the United States with 114 systems. Japan ranks third with 30 systems. There are 18 supercomputers from Top500 in France, and 16 in Germany. In terms of total performance, the United States ranks first with 644 petaflops. The total performance of Chinese supercomputers is 565 petaflops, Japanese - 530.^[2]

2019

China dominates the number of supercomputers in the top 500

In 2019, the growing share of Chinese supercomputers is indicative - their 45.6% of the total (228 out of 500). For comparison, the share of American - 23.4%.

Sberbank supercomputer was among the 30 most powerful in the world

In November 2019, an updated rating of the most powerful supercomputers in the world was published. Christofari took 29th place in it.

In addition to this system, the rating includes two more cars from Russia: Lomonosov-2 from Moscow State University and SuperComputer of the main computing center of Roshydromet, which took 107th and 465th places, respectively. In June 2019, they were in higher positions - in the 93rd and 364th.

An updated rating of the most powerful supercomputers in the world has been published

According to RBC, citing Dmitry Konyagin, head of professional products at Nvidia, Christofari is 2.7 times faster than the previous fastest supercomputer in Russia, based on average power during tests. From December 12, third-party companies will be able to use Sberbank's supercomputer.

The ranking of the top 500 has been compiled since 1993 twice a year by scientists from the USA and Germany. The minimum threshold for getting into it this year reached 1.022 petaflops.

The systems on the list are built according to the maximum performance in the LINPACK test, for Christofari it is 6.669 petaflops. For comparison, the maximum performance of the Lomonosov-2 supercomputer is 2.478 petaflops.

The first place on the list for several years has been occupied by a system Summit built in the laboratory power engineering specialists of the US Department of Commerce in Oak Ridge. Its maximum figure in LINPACK today is 148.6 petaflops.

The leaders in the number of supercomputers represented are China (228 cars), followed by the United States (117). Japan is in third place (29 supercomputers).

There are more and more Chinese supercomputers in the ranking, their performance is also increasing. By October 2019, supercomputers in the United States account for 37.1% of total power, while Chinese ones account for about 32.3%.^[3]

2017

Reasons for Chinese dominance in HPC

The new edition of the list of the most powerful supercomputers Top500 does not accurately reflect the real state of affairs in the HPC industry. China's current hegemony is nothing more than a consequence of brief stagnation in the US and Europe.

The Top500 website provides all possible statistical information regarding the world leaders of HPC, but without interpretation, which opens up the possibility of playing numbers. With this in mind, let's try to analyze some features of the latest edition of the Top500 list.

The SC 17 supercomputer conference, held in Denver in November 2017, fell on a kind of period of temporary lull. Leading US companies have gone into the shadows for various reasons. At the same time, they are preparing for a significant event that will happen after 2020. At this time, the creation of a long-awaited exaflop computer is expected. In preparation for it, already in 2018, they will present several computers approaching exaflops, the so-called pre-exascale computer.

Japan and a united Europe are striving for the same cherished goal.

Japan The company Fujitsu intends to build a new exaflop version of the K-computer by 2021, but this time not on SPARC-64 processors traditional for this company, but on processors. On the ARM same ARM processors, the Cavium ThunderX2 version creates an exaflop computer according to the European project Mont-Blanc. The development is supervised by the Supercomputer Center in Barcelona, ​ ​ and it will be built by a French company, Atos which includes a well-known server manufacturer. Bull

Chinese phenomenon and its causes

Meanwhile, SC 17 is dominated by China and this phenomenon deserves the most serious attention. Until a few years ago, Chinese supercomputers seemed like paper dragons. In 2017, they are at the top of their fame and it turns out that by 2021, China should also expect the appearance of three computers with performance approaching exaflops, and on its own processors.

China has long established the production of processors by reverse engineering methods, relying on well-known architectures. Moreover, Alpha, MIPS and SPARC have already been abandoned there, in 2017 in the focus of x86 and Power8. It should not be surprised that the most active processes taking place in the Chinese HPC segment were able to cause unexpected changes in the Top500. How to understand it - is it a trend or an outlier?

On November 13, 2017, SC 17 presented another list of Top500 supercomputers, different from those preceding the obvious transition of leadership to China. This country not only took the first two places, but also bypassed the United States in the total number of computers on this list. Climbing the pedestal was surprisingly rapid, because back in June 2017 the ratio was 169:159 in favor of the Americans and suddenly the Chinese led with a score of 202:144. It would seem that dozens of supercomputers arose from nowhere. What happened? Has the Celestial Empire really become the "number 1" in HPC?

Intuition imposes doubts, it makes you think about the possible cunning of numbers and about the ability of those who want to deftly use the fairly outdated method of forming the famous list, which turned exactly a quarter of a century in 2017. During this time, 50 editions of the Top500 were published. And all these years, the drafting process has been constantly led by Jack Dongarra, one of the creators of the LINPAC test system. The rating system he created allows you to express the computer performance by the number of floating point operations performed per second, briefly in flops. For the fastest, add "peta." Petaflops is equal to 1 trillion operations per second.

For all members of the HPC community, Dongarra has become an absolutely iconic figure, he is the main character of all supercomputing events in the world and we should be grateful to him and his colleagues for their efforts to create LINPAC and Top500, which have become industry crystallization centers. Nevertheless, everything is aging, LINPAC and Top500 are no exception. They, like any system, accumulate errors during their lifetime, which opens up the possibility of using tests for other purposes, as clearly evidenced by the visible success of China. The LINPAC tests and Top500 compilation methodology were created in conditions radically different from the current one.

Mismatch of the list with the real position is possible for two reasons. The first is that in some cases systems are tested that do not actually serve the purposes of HPC, but are nevertheless capable of performing tests. As a result, "the wrong ones" are included in the list. Conversely, for a number of reasons, "those that are necessary" systems that are really intended for HPC purposes do not seek to get into the number of tested ones. As a result, there was an obvious mismatch between the real fleet of computers serving for HPC and the list of Top500.

The discrepancy became possible because recently systems have appeared outwardly surprisingly similar to HPC, but in fact they are not. The need to work with clouds and large amounts of data, solving machine learning problems and other loads characteristic of the current time have led to the emergence of systems like HPC. They use almost the same servers and DSS, almost the same network tools. They have qualities such as horizontal scaling, high level of parallelism, the use of GPU in addition to CPU, low power consumption, dynamic resource management, etc.

And yet, despite the similarity, such systems are more likely to be attributed to the corporate class, sometimes they are called hyperskelers. The owners of hyperskelers may be large online trading companies or providers of other information services. The emergence of the described systems expands the market for technologies used in HPC, with all the ensuing positive consequences.

The proximity of hyperscalers to HPC creates the temptation to aim them at solving tasks for which they are not intended at all, with the artificial goal of getting into the Top500. What is the reason for such a desire? It is difficult to say from the outside, but in many cases a variety of economic or political and far from scientific reasons are not excluded. The compilers of the Top500 cannot be blamed for this, they use the technical data available to them, and the issues of financing and ops budgets (black ops budgets), characteristic of the supercomputer world, are beyond their competence.

The following observation can be used as evidence of the above statement. The efficiency of consolidated operation of clusters on large loads, for which HPC is created, is largely determined by the quality of interconnections. Nevertheless, in recent years, the number of installations has been growing noticeably, where cheaper Ethernet is used as an interconnect, and not Infiniband.

The ratio of specialized and non-specialized systems is shown in the figure below, their number has almost equalized and there is a noticeable trend towards the noted trend.

It should be noted that it is not only hyperscalers that borrow ideas from HPC. Progress is a two-way street, for example, clouds and virtualization received from corporate HPC systems.

Oddly enough, along with those who want to mimic and pass off their systems as HPC and get on the coveted list, there are also such HPC owners who do not want to test their systems. In some countries, such computers can be located in closed cities, in others - even in universities. Apparently, one of the world's most powerful Blue Waters supercomputers at the University of Illinois is not rated LINPAC. The owners of the world's largest hyperscalers, the same Google or Amazon, do not strive for fame either . If they put together their computing power, they would get a total of thousands (!) Of exaflops.

New metric

As a result, according to the Hamburg account, real supercomputers can be called those that fall into the first hundred, and of the remaining ones, those that were once at the top, but lost the top positions during their life.

Who needs computers designed for the industry to earn their place in the Top500 can only guess. Either this is a struggle for national prestige, or an excuse for using undisclosed sources of funding, or something else. The conclusion is perhaps one - when rule violations become a system, then you need to change the rules. As it stands, Top500 has ceased to reflect the true state of affairs of the HPC. Therefore, you should not wait for the Top500 to mark its 100th edition.

One of the attempts to create an alternative assessment system at HPC was the creation of the Virtual Institute for I/O in 2016.

The result of this public organization should be a list of IO500 that will more accurately correspond to modern ideas about high productivity.

European response

The united Europe (EU), which seems tired of staying away from the main events, is preparing its response to the challenge posed by China, along with the United States. It came to the point that in the top ten of the current Top500 there is not a single computer from the EU. In third place is the Swiss Piz Daint, but this country is not a member of the European Union.

Among several pan-European projects aimed at restoring positions are EXDCI (European Extreme Data and Computing Initiative), ETP4HPC (European Technology Platform for HPC) and Partnership for Advanced Computing in Europe (PRACE).

Europe intends to spend more than a billion euros in the coming years on the construction of an exaflop computer. In the period from 2021 to 2022, two pre-exascale computers will be built, and in 2023 two computers with performance over exaflops. All of them will be based on 64-bit ARMv8 processors.

But do not forget that Europe is a country of high scientific culture and truly breakthrough technologies may appear here. For example, the quantum computer Atos QLM. According to the Franco-Belgian group of scientists that created it, it is a self-learning computer (Atos Quantum Learning Machine).

The computer is compact, its height is about half of human height. To solve problems, Atos has developed a quantum modeling system based on the universal programming language aQasm (Atos Quantum Assembly Language). The description of the system is obscure to the public, from a user point of view, much more informative reviews from specialists from Oak Ridge National Laboratories, which acquired this product.

For the user, Atos looks like many Linux servers with a rather peculiar file system. They run an application for modeling quantum circuits, which is a cocktail of vector multiplications based on linear algebra with a limited set of data. You have to work on it using the "brute force" method, you have to write programs in the form of text files like the assembly language, where each line is a command for one of the modules (qubits), in total there can be from 30 to 40. This file is translated into a form that Atos understands and executed. The execution result is returned as a printout of the new qubit state. While programs are relatively simple, there are no special difficulties, but they can become more complicated due to the theoretical inability to place intermediate checkpoints during the execution of the program

US outlook for 2018

Summit and Sierra supercomputers, which IBM will build on the basis of Power 9 processors with the participation of Nvidia and Mellanox, and Aurora, a joint development of Intel and Cray, are competing for prizes in 2018 . All are for U.S. national nuclear centers.

China outperforms the United States in the number of the most powerful supercomputers

China was ahead of the United States in the number of most powerful computing systems. From the Top500 rating published in November 2017, which includes the most productive supercomputers in the world, it became known that there are 202 such systems in the PRC, and only 143 in the United States.^[4]

Sunway TaihuLight is the fastest supercomputer in the world

Experts note that China has achieved the highest result in the 25-year history of the Top500 rating, and the US indicator, on the contrary, has become the worst in the last quarter of a century. Six months ago, in June 2017, the PRC and the United States were on the list in reverse order: there were 160 and 169 most powerful computing systems in the countries.

It is also reported that Japan, where the 35 most high-performance supercomputers in the world are located, has become the third in the list. This is followed by Germany (20 systems), France (18) and the United Kingdom (15).

The list of the most powerful computing systems used for research in the field of climatic and weather changes, oil exploration, nuclear power, DNA sequencing, biomolecule modeling, etc., as in June 2017, was headed by the Chinese supercomputer Sunway TaihuLight with a capacity of 93 petaflops (quadrillion floating point operations per second). The top 5 also includes the Tianhe-2 system (33.86 34 petaflops), created by specialists from the Chinese National University of Defense Technology (NUDT), the Swiss supercomputer Piz Daint (19.59 petaflops), the Japanese system Gyoukou (ZettaScaler-2.2) (19.14 petaflops) and the American Titan (17.4 petaflops), which ranked higher in June 2017.

Rating of Top500 for November 2017

The fastest Russian supercomputer installed at Moscow State University (MSU) is in 63rd place in the November ranking, although six months ago this system with a capacity of 2.10 petaflops was located in 59th place.^[5]

The fall of Lomonosov-2 in the ranking of the most powerful supercomputers

On June 19, 2017, an updated rating of the most powerful TOP500 supercomputers was published. The Chinese system remained the leader, and the fastest supercomputer in Russia in a year slipped by 18 places at once.

Sunway TaihuLight retained the first place in the list of the most productive supercomputers, the performance of which by June 19, 2017 is 93 petaflops (quadrillion floating point operations per second) with a theoretically possible increase to 125 petaflops.

Lomonosov-2 supercomputer

The second place in the rating was taken by another Chinese system - Tianhe-2. Here the speed is declared at the level of 33.86 petaflops and the maximum possible indicator of 54.9 petaflops.

The top three were the Swiss supercomputer Piz Daint, which has current and potential performance of 19.59 and 25.33 petaflops, respectively. It is curious that for the first time in a long time in the top 3 there was no complex from the United States. The most powerful among such - Titan - is located on the fourth line with a result of 17.59 petaflops.

At the same time, most of the TOP500 supercomputers are based in the United States (169 positions in the ranking). In second place is China with 160 systems, in third - Japan (33), in fourth - Germany (28). France and Great Britain each took 17 positions on the list.

The fastest Russian supercomputer is still Lomonosov-2, installed at Moscow State University (MSU). Its output is 2.10 petaflops and its peak output reaches 2.96 petaflops. Such values ​ ​ allowed the system to be located in 59th place, while in June 2016 it occupied 41st position. In total, only three supercomputers from Russia remained in the ranking: in addition to Lomonosov-2, these are Lomonosov and RSK Tornado, which are located at Moscow State University. Lomonosov and St. Petersburg State Polytechnic University, respectively.

Most supercomputers in the ranking - 464 out of 500 - run processors. Intel Accelerators are included in supercomputer configuration 91, and most often GPUs act as this. Nvidia^[6]

2016

SC Bull sequana entered the Top500 rating

On December 7, 2016, the company Atos and the French to power France Alternative Energies and Atomic Energy Commission (CEA) announced that they were included in the list of Top500 most powerful in the supercomputers world. supercomputer Bull sequana

The Top500 system included in the rating is one section on the Bull sequana supercomputer platform, which has 85 computing blade modules connected to each other through BXI technology. The supercomputer supports 220 Intel Xeon Phi 7250 processors, and its peak theoretical performance ~ 670 teraflops. The performance in the Linpack test ~ 380.5 teraflops.

Chinese system on Chinese processors led Top500

On June 20, 2016, a new edition of the rating of the most powerful supercomputers in Top500 was announced. The first place went to the Chinese computer system Sunway TaihuLight, which uses processors developed domestically.

The Sunway TaihuLight has a performance of 93 petaflops (10 in the 15th degree of floating point computing operations per second), which is twice as fast and three times more efficient than the computer of Tianhe-2, the leader of the previous rating based on Intel chips.

Sunway TaihuLight, used for climate modeling and biomedical research, contains 10.65 million nuclei - about 41 thousand nodes. The supercomputer was developed by the National Research Center for Parallel Computing and Technology of the PRC, it was placed at the National Supercomputer Center in Wuxi, Jiangsu Province.

Sunway TaihuLight Supercomputer Placement View

Sunway TaihuLight - Sunway MPP, Sunway SW26010 260C 1.45GHz, Sunway

Sunway TaihuLight became the first Chinese supercomputer in history to lead the Top500 and devoid of American chips. The system has processors created at the state-owned Shanghai High Performance IC Design Center.

In addition, China overtook the United States for the first time in the number of supercomputers on the Top500 list. By June 2016, there are 167 such computing systems in the PRC against 165 in the United States.

Jack Dongarra, a specialist in computer and system theory at the University of Tennessee, believes that the United States will not be able to create a supercomputer that will be comparable in power to the Sunway TaihuLight until 2018, especially since the Chinese system has a theoretical maximum performance of 125.4 petaflops.

China has taken a strong lead in this race, adds Horst Simon, deputy director of the Lawrence Berkeley National Laboratory (one of the main users of supercomputers who helps form the list of Top500)

In the second position of the updated rating is the Chinese Tianhe-2, which has a performance of 33.9 petaflops. The third was the American Titan - Cray XK7 (17.6 petaflops) from Cray Inc. The Russian supercomputer Lomonosov-2 (2.1 petaflops) took 41st place in this list, which is 10 places worse compared to the previous edition of the rating.^[7]

2015

The success of Russian ICs is two in the first hundred

On November 26, 2015, the Tor500 project published a list of the highest-performing computers in the world^[8].

The fastest Russian supercomputer on the list is Lomonosov-2, it is installed in. MSU The performance of the system is 1,849 PFLOPS, which corresponds to 35 positions in the list. The top 100 also includes a supercomputer ("Lomonosov" 94th place). In this period Russia , seven systems are represented in the list of Tor500.

At the head of the list of the most productive ICs is a Chinese Tianhe-2. The product showed a result of 33.86 PFLOPS (¹⁰¹⁵ floating point calculations per second) in the Linpack test.

The number of Chinese computers on the list has almost tripled, and systems from the United States - the smallest in the entire existence of the project Top500 since 1993 - 200. Six months ago there were 231 of them. There are fewer European systems - 108 against 141. The number of systems installed in the Asian region has increased - from 107 to 173, while 109 - in China. Japan is represented by 36 systems (in the previous Top500 - 40).

In the second position of the list is the Titan system (Cray XK7), USA. Its performance indicator is 17.59 PFLOPS.

The top ten included two fresh systems: Trinity (Cray XC, 8.1 PFLOPS, USA) and Hazel-Hen (Cray XC, 5.6 PFLOPS, Germany) in sixth and eighth positions, respectively.

Six of the ten most productive systems were commissioned in 2011 or 2012, Tianhe-2 in 2013, Trinity, Hazel-Hen and Shaheen II (Saudi Arabia) in 2015.

RSK products represent 50% of Russian ICs in Tor500

On July 24, 2015, it became known that the share of the RSK group of companies increased by 50% among Russian systems in the list of Top500 most powerful supercomputers in the world.

The Top500 rating (as of July 2015) includes 4 supercomputer complexes manufactured by RSK, their total peak performance exceeds 2 PFLOPS. In total, the current version of the list contains 8 supercomputers from Russia.

Two computing systems RSK were created and installed as part of the project of one of the most modern and large computing centers in Russia for a Peter the Great St. Petersburg Polytechnic University (SPbPU) total peak performance of more than 1.1 PFLOPS.

The RSK Tornado Polytechnic cluster system ranks 107th with a capacity of 658 TFLOPS on the LINPACK test. The peak performance of this SPbPU computing cluster based on the RSC Tornado architecture with direct liquid cooling is 829 TFLOPS.

The second part of the SPbPU hybrid computing complex consists of a unique superdense massively parallel PetaStream RSC system with direct liquid cooling and a peak capacity of 258 TFLOPS.

Supercomputer "Polytechnic RSC PetaStream" demonstrated the performance of 170.5 TFLOPS on the LINPACK test, taking 466th position in the current version of the Top500 rating. This SPbPU computer system, developed and manufactured by RSK specialists in Russia, is built on the basis of 60-core Intel Xeon Phi 5120D and Intel Xeon E5-2600 v2 processors, Intel server boards and Intel SSD DC S3500 solid-state drives for corporate data centers.

At the 174th position of the list, the supercomputer MVS-10P with a peak performance of 523 TFLOPS, installed by RSK specialists at the Interdepartmental Supercomputer Center of the Russian Academy of Sciences (MSC RAS).

The 244th place in the Top500 is occupied by the computing cluster "RSC Tornado" of SUSU "with a peak capacity of 473 TFLOPS, operated since 2009 at South Ural State University (SUSU) in Chelyabinsk.

2014: Russian supercomputers in the world Top 500 ranking

On November 18, 2014, it became known about the strengthening of positions Russia in the new final edition of the World Top 500 rating supercomputers.

Compared to the Top 500 six months ago, the number of Russian-made HPC class systems increased from five to nine. The highest achievement of the Russian company T-Platforms is the 22nd line of the rating with the new system - A-Class.

In the new edition of the list, the first place is occupied by the Chinese cluster "Milky Way 2" (Tianhe-2) (for the fourth time in a row) with a peak performance of 54.9 Pflops and 33.9 Pflops performance on Linpack.

The Lomonosov cluster, a product of T-Platforms, dropped 16 positions compared to the previous Top 500 rating and took 58th line with indicators of 1.7 and 0.9 Pflops, respectively.

RSK products in the rating are two new systems: the first in 81 rating positions with results of 829 and 658 Tflops, the second in 390 positions with results of 573 and 171 Tflops. The systems are presented in various RSK architecture - Tornado and Petastream.

The systems were developed as part of the project to create a new supercomputer center at St. Petersburg State Polytechnic University. The RSK Tornado Polytechnic cluster system based on the RSK Tornado architecture with direct liquid cooling and the latest Intel Xeon E5-2697 v3 server processor was included in the top hundred of the Top500 list, ranking 81st.

The second part of the new hybrid computing complex SPbPU consists of a unique superdense massively parallel PetaStream RSC system with direct liquid cooling and a peak capacity of 258 TFLOPS. Supercomputer "Polytechnic RSC PetaStream" demonstrated the performance of 170.5 TFLOPS on the LINPACK test, taking 390th position in the current version of the Top500 rating. This SPbPU computer system, developed and manufactured by RSK specialists in Russia, is built on the basis of 60-core Intel Xeon Phi 5120D and Intel Xeon E5-2600 v2 processors, as well as Intel server boards and Intel SSD DC S3700 for corporate data centers.

The fall 2014 ranking Top500 includes two more old RSK computing systems based on the RSC Tornado cluster architecture with liquid cooling, Intel Xeon server processors, Intel server boards and Intel SSDs.

At 133 of the list is the MVS-10P supercomputer with a peak performance of 523 TFLOPS, installed two years ago at the Interdepartmental Supercomputer Center of the Russian Academy of Sciences (JSCC RAS). 190th place in the Top500 is occupied by the computing cluster "RSK Tornado SUSU" with a peak capacity of 473 TFLOPS, operated since 2009 at the South Ural State University (SUSU) in Chelyabinsk.

The Russian newcomer Top 500 - a cluster of the Volga Scientific and Educational Center for Supercomputing Technologies, which is one of the main training units of Nizhny Novgorod State University (NNSU), Lobachevsky, took 189th place with the results of 349 and 290 Tflops. Only a fragment of "Lobachevsky" entered the Top 500.

In the Russian Top 50, the cluster is represented in two positions: 6th place (part with Nvidia accelerators) and 48th place (part with Intel Phi coprocessors).

Two more systems based in Russia, participants in the new Top 500 - two Hewlett-Packard systems from unnamed IT service companies. One system appeared in the ranking earlier, now it took 457 place with indicators of 317 and 161 Tflops. The second is declared for the first time and takes 337 place with a score of 295 and 189 Tflops.

In the previous ranking Russia , five computers were presented, in the current nine.

2012:500 largest supercomputers

The ranking of the 500 largest supercomputers, published on November 12, 2012, confirmed that the world's highest-performing supercomputer operates at Oakridge National Laboratory (Oak Ridge National Laboratory, USA).

Titan supercomputer (Cray XK7 system), created by Cray using Nvidia Tesla K20 accelerators and running on 560,640 processor cores, of which 261,632 are NVIDIA K20x accelerators, in the Linpack test shows a performance of 17.59 Petaflops/s (quadrillion operations per second). The new rating was announced at the largest international exhibition on high-performance SC12 computing, which is taking place these days in Salt Lake City, USA.

Thus, Titan removed from the position of the leader the Sequoia BlueGene/Q supercomputer deployed at the Livermolsk National Laboratory (Lawrence Livermore National Laboratory, USA), which in June 2012 showed 16.32 Petaflops/s in the Linpack test, while operating on more cores - 1,572,864.

The top five supercomputers also included:

• Fujitsu K RIKEN Advanced Institute for Computational Science (AICS) in Kobe, Japan (10.51 Petaflops/s),
• Mira BlueGene/Q at Argonne National Laboratory, USA, (8.16 Petaflops/s) and
• JUQUEEN BlueGene/Q at the Forschungszentrum Juelich Germany (4.14 Petaflops/s).

Half of all supercomputers included in the Top 500 rating are installed in the United States, 105 are in Europe and 124 in Asia.

Interestingly, the Russian supercomputer with a capacity of 520 Tflops of the JSCC RAS supercomputer center also entered the world ranking of the top 500 most powerful computing systems in the world. Recall that a year ago, the most powerful was the Chinese supercomputer.

23 supercomputers included in the published rating have a performance above 1 Petaflops/s, 62 systems use accelerator and coprocessor technologies, 84% have processors running on 6 or more cores, 46.2% - on 8 or more cores.

Intel remains the leader in the supply of processors for supercomputers - its processors are used in 76% of systems, followed by the AMD Opteron family (12%), IBM Power processors work in 10.6% of super systems. Communication in most supercomputers is provided by InfiniBand technology, and Gigabit Ethernet is gradually losing ground.

Notes