Gigaflops processors. A supercomputer with petaflops performance is just around the corner

Studies show that, on average, desktop computing power lags behind supercomputer performance by 13 years. In other words, today's professional PCs are nearly as powerful as 13-year-old supercomputers in terms of performance. This is why HPC market research is - good way to assess the direction of development of mass computers of the future. Not so long ago, supercomputers surpassed the performance bar of one teraflops (trillion floating-point operations per second), and they are not far off reaching the performance level of petaflops (quadrillion flops, or 1015 floating-point operations per second). , while tera-calculations will remain with the average PC user ...

American professor and writer Steve Chen tried to imagine what level of performance will be sufficient for solving various problems in the future. In his opinion, for aerodynamic problems, a performance of several petaflops is enough, for molecular dynamics problems it will take 20 petaflops, for computational cosmology - fantastic performance at the level of 10 exaflops (one exaflops is equal to a quintillion, or 1018 flops), and for problems of computational chemistry, more more powerful processors. Steve Pavlovsky, Intel Senior Research Engineer and Chief Technology Officer and General Manager of Architecture and Planning for Intel's Digital Enterprise Group, believes computers with a sextillion performance, or 1,021 floating point operations per second, will be available by 2029.

Steve Pavlovsky believes that the challenges and achievements of today's supercomputers will be the challenges and achievements of tomorrow's desktop PCs. The HPC market is growing - its volume has already reached $ 10 billion, and in some sectors the annual sales growth exceeds 30%; the number of high-performance professional computers based on Intel processors sold worldwide is also growing.

Just 60 years ago, the ENIAC tube computer, considered the technological pinnacle of high performance computing, had just 20 cells. random access memory... In the mid-60s, the CDC 6600 supercomputer appeared, with a performance reaching 9 megaflops. And only in 1997, the ASCII Red supercomputer, containing 9298 processors Intel Pentium Pro, has reached a performance level equal to teraflops. Today, a system based on 464 quad-core processors Intel Xeon the 5300 series, in a much smaller footprint, has six times the peak performance.

When will the performance of the petaflops level (that is, thousands of teraflops) be reached or, as Steve Pavlovsky figuratively puts it, will the “sound barrier” of the pet performance be overcome? And when will peta computing become basic to ordinary computer systems?

It is estimated that the first peta supercomputers will appear in 2008-2009 - to determine this timeline, it is enough to take the performance parameters of the world's fastest computers published at www.top500.org and extrapolate them in accordance with the observed growth trends. However, in order to create peta computers for the mass market, there are many serious problems to be solved. To this end, Intel, together with partners, is conducting research in the following areas:

• performance;
• throughput memory;
• interconnects;
• power management;
• reliability.

According to Steve Pavlovski, to achieve the level of pet computation using modern technologies increasing the performance of semiconductor microcircuits will require the creation of a processor with 100 thousand computational cores. For the practical implementation of such systems, it will be necessary to significantly increase the density of the arrangement of nuclei on the crystal. Today, there is a fierce debate over the architecture of future computers - which is better: many small cores optimized to speed up parallel computing, or several larger cores designed to speed up sequential computing? Leaning towards the first path of development, researchers understand that they are setting themselves the laborious task of transferring the software industry to the parallel programming rails ...

Another area of ​​Intel research is the organization of connections between computational cores. Bus connections take up less space, have high bandwidth, and scale well, but are energy inefficient. The second option is ringing the cores for signaling, the disadvantage of which is the low level of scalability as the number of cores increases. The third option is a matrix architecture, when each core communicates with each through a chain of neighboring cores.

It is worth remembering that at the fall Intel Developer Forum (IDF) in San Francisco, a prototype of an 80-core processor was unveiled that could potentially provide teraflops performance for desktop computers... According to Intel's chief technology officer, Justin Rattner, the estimated release date for such a processor on the market is 2010 or even earlier. The processor prototype is based on the x86 architecture and Intel developments such as high performance computing on a chip (HPC-on-chip), a new structure of memory elements connections, new energy-saving technologies, etc.

In 2006 Intel Corporation announced global program research, called Tera-Scale Computing, and brings together more than 80 different research projects around the world, distributed in three main areas: improving technologies for the design and manufacture of silicon crystals, platform optimization and new approaches to programming. In his speech at IDF, Justin Rattner noted that the necessary steps towards the tera era will be taken over the next decade. For example, modern research is aimed at optimizing the operation of cache memory, its configurability depending on the tasks to be solved, and at developing parallelism in accessing multiple cores to shared memory. Intel also plans to integrate a broadband digital self-tuning wireless transceiver into its dies, and applications based on integrated silicon photonics are just around the corner.

“High data transfer rates between compute cores and memory are an important issue,” stresses Pavlovsky. - The memory must have an extremely high bandwidth. At the same time, if we increase the clock frequency of the memory channel, then soon enough we will face the physical limitations that copper conductors impose. " One of the possible ways to overcome these limitations is to increase the number of memory channels, but this increases the size of the processor and its cost. “We will have to look for more exotic data transmission technologies,” says Pavlovsky. "We estimate that peta processors will require about 500GB / s of memory to run."

The next most important aspect of the operation of pet computers is the speed of the I / O system. Intel scientists are now working to deliver data transfer rates of up to hundreds of gigabytes per second (GB / s).

Yet the biggest challenges in peta devices are power supply and reliability. The power consumption of a modern large data processing center (DPC) averages 9-10 MW. The power consumed by a computer with 100 thousand cores can be about 20 MW. To this must be added the power required to cool the pet computers. At the current cost of electricity, the cost of supplying the peta system alone would exceed $ 14.6 million per year. That is why the question effective use electricity is extremely important, which dictates the use of energy-saving technologies at all levels - from transistors to data centers:

• at the transistor level - strained silicon technologies, technologies for reducing leakage currents, etc .;
• at the processor level - load balancing based on multithreading;
• at the system level - high-precision power consumption control depending on the system load;
• at the data center level - the use of advanced liquid and air cooling systems, as well as vertical integration of heat dissipation solutions.

Moreover, researchers predict the emergence of completely unexpected problems associated with ... cosmic rays. After all, peta processors with highly integrated computing elements will use transistors so small that they will be influenced by energetic particles that make up cosmic rays and can cause random data failure when they enter the transistor. As the density of transistors on a chip increases, the number of such random failures will grow rapidly. "If the number of cores on a chip reaches 100,000, such failures will become unmanageable," says Pavlovsky. - They will have an increasing influence on the operation of the system, and they will need to be fought. We have already started research in this direction. " Advanced technologies Reliability assurances include the use of parity and error correction codes, and the use of redundant cores to validate the computational results of the main system cores.

Computing technology is developing by leaps and bounds. Therefore, it is likely that at the moment when this article was published the light saw the light of a new "monster of computing". We would like to introduce you to the top ten leaders for November 2012.

1.Titan (USA) - 17.59 petaflops

The first place was taken by the American supercomputer Titan, created with the participation of Cray and Nvidia. It is located at Oak Ridge National Laboratory in Tennessee, which is owned by the US Department of Energy. The Titan can perform 17.59 quadrillion floating point operations per second, which is equivalent to 17.59 petaflops.

Titan consists of 18688 knots. It is built on a hybrid architecture: each supercomputer node includes a 16-core AMD Opteron processor and an Nvidia Tesla K20X graphics accelerator. The use of GPUs can reduce the power consumption of the system.

Titan is being used to design energy efficient engines for vehicles, to simulate the effects of climate change and to study biofuels. Oak Ridge leases the supercomputer to other research organizations.

2. Sequoia (US) - 16.32 petaflops

The Sequoia supercomputer, also owned by the US Department of Energy, runs on 1,572,864 cores. Sequoia is being developed by IBM for the National Nuclear Safety Administration as part of its advanced computing and simulation program.

Sequoia will be used primarily for simulating nuclear explosions, replacing the ASC Purple and Blue Gene / L supercomputers at Livermore National Laboratory. Sequoia will also be able to solve problems for the needs of astronomy, energy, the study of the human genome and climate change.

Sequoia is built on the Blue Gene / Q architecture, the latest generation in the Blue Gene line of supercomputing architectures. The supercomputer consists of 98,304 computing nodes and has 1.6 PB of memory in 96 racks, located on an area of ​​300 square meters. m. Used 16 or 8-core central processing units Power Architecture, manufactured according to the 45 nm process technology.

IBM has created a computer that can solve 20 quadrillion of various mathematical operations in one second. This means that if 7 billion people took calculators and began to do mathematical calculations at the same time without respite, all 24 hours a day, all 365 days, then these operations would take up to 320 years, no less. But now you don’t need to do that, because Sequoia was born. The computer will carry out such calculations in just an hour.

3.K computer (Japan) - 10.51 petaflops

K computer is a Japanese supercomputer manufactured by Fujitsu, launched in 2011 at the RIKEN Institute for Physical and Chemical Research in Kobe. The name comes from the Japanese prefix "kei", meaning 10 quadrillion and at the same time designating the capital, that is, an allusion to the "main computer"

As of June 2011, the system had 68,544 8-core SPARC64 VIIIfx processors housed in 672 computing racks, representing 548,352 computing cores manufactured by Fujitsu using 45nm process technology. The supercomputer uses water cooling to reduce power consumption and increase packaging density.

4. Mira (USA) - 8.16 petaflops

Using the supercomputer IBM Blue Gene / Q (Mira), American scientists will try to model the universe. Scientists hope to get answers to the most interesting questions about the origin of the universe. The computer is supposed to simulate and consistently calculate the 12 billion years that have passed since the Big Bang.

A supercomputer consists of 50 thousand computational nodes, each containing 16 cores. The computer uses a massive 70 petabyte storage and liquid cooling system. Mira is capable of 8 quadrillion operations per second.

5. JuQueen (Germany) - 5.9 petaflops

JuQueen, the most powerful supercomputer in Europe, has been officially launched in the German city of Julich (North Rhine-Westphalia). Its performance is 5.9 petaflops or 5.9 thousand trillion operations per second.

JuQueen processors have a total of almost 459 thousand cores. At the same time, they were developed using energy-saving technologies. The system will be cooled using circulating streams of water with a temperature of 18 degrees. Experts point out that this machine is about 100 thousand times more powerful than the most modern personal computer.

The computer was developed by the IBM corporation. The project was financed from the funds of the largest scientific organization of the Federal Republic of Germany - the Helmholtz Center, the federal budget, as well as from the treasury of North Rhine-Westphalia. The exact amount was not disclosed.

6.SuperMUC (Germany) - 2.9 petaflops

SuperMUC, the second most powerful supercomputer in Europe, was launched at the end of June 2012. The supercomputer was created to solve complex scientific problems in the physics and dynamics of fluids. The machine runs on SUSE Linux Enterprise Server platform. The SuperMUC on the System X iDataPlex from IBM is equipped with more than 155,000 processor cores for a combined maximum performance of about 3 petaflops.

A special feature of SuperMUC is an innovative technology for cooling the system with warm water, developed by IBM, which is based on the blood circulation system in the human body. As a result, SuperMUC spends 40% less energy on cooling systems than "classic" data centers, and also allows you to accumulate and use the saved energy to heat the buildings of the Leibniz Computer Center.

7. Stampede (USA) - 2.7 petaflops

The Texas Advanced Computing Center (TACC) at the University of Texas has created a supercomputer capable of 2.7 quadrillion floating point operations per second. TACC is part of the XSEDE (Science and Engineering Advanced Discovery Environment) project, which aims to provide researchers with access to supercomputing resources.

At the heart of Stampede is Dell's hyperscalar architecture using Intel Xeon E5-2680 8-core processors. Xeon processors deliver over 2 petaflops of performance. Work on the project is still pending, and in 2013 Stampede will also use new Intel Xeon Phi coprocessors designed to perform parallel computations, which will be responsible for more than 7 petaflops of system performance. This will increase the total system performance up to 10 petaflops.

In addition to Xeon Phi, the supercomputer will use 128 next-generation GPUs from NVIDIA to provide remote virtualization. System performance can grow up to 15 petaflops as processors are installed Intel new generations. Mellanox is another supplier of components for Stampede, providing 56Gbps Infiniband networking equipment.

The supercomputer's cooling system is built on the principle of isolating hot zones and involves the use of built-in cooling modules, which allows placing equipment with a high density of up to 40 kW per rack. The power distribution system supplies 415V to the racks and 240V to the servers. The power requirements of the Stampede and Ranger systems are supplied by a 10MW power substation.

8.Tianhe-1A (China) - 2.57 petaflops

Tianhe-1A is a supercomputer designed National University defense technologies of the People's Republic of China. The computational speed of the supercomputer is 2.57 petaflops.

The Tianhe-1A uses 7168 Nvidia Tesla M2050 GPUs and 14336 Intel Xeon server processors. According to Nvidia, the supercomputer uses electrical energy three times more efficiently than other electronic computers in its class. A supercomputer built exclusively on the basis of central processing units(CPU), with a comparable computation speed, would consume more than 12 MW of electrical energy. The electric power consumed by Tianhe-1A is 4.04 MW. Without GPUs, a supercomputer of comparable performance would require more than 50,000 CPUs to be installed.

The construction of the supercomputer cost $ 88 million, and the annual operating costs are about $ 20 million. The maintenance employs about 200 specialists. The main area of ​​work is research on oil production and aerodynamics. Declares “ open access»To a supercomputer, which theoretically allows its use by other countries.

9.Fermi (Italy) - 1.7 petaflops

Fermi is in ninth place. Systempostedon the servers of the non-profit consortium Cineca, which includes 54 Italian universities and research organizations.Fermi consists of 10,240 1.6 GHz PowerA2 processors, each with 16 cores. In total, the computer has 163,840 computing cores.Each processor comes with 16GByte RAM (1GByte per core).Fermi is used by Italian and European research teams to perform the calculations needed in large-scale research projects to solve fundamental problems in science and technology.The system is named after Enrico Fermi, an Italian nuclear physicist.

10.DARPA Trial Subset (USA) - 1.5 petaflops

This system is an IBM Power 775 server with 63360 cores, which achieves a performance of 1.5 petaflops. Other information on this moment no.

In conclusion…

Russian development - the supercomputer "Lomonosov", owned by the Moscow State University named after M.V. Lomonosov, in this list (at the end of 2012) it takes twenty-second place. Its performance was 0.9 petaflops. The main reason why domestic cars do not occupy leading positions in international ratings, Russian manufacturers unanimously cite the lack of adequate funding.

The main type of nodes providing over 90% of the performance of a supercomputer is T-Blade2. This supercomputer platform was created by T-Platforms engineers from scratch - all its boards and mechanical components are the company's own patented developments. In terms of computational density per square meter of area, T-Blade2 has no analogues in the world. So Russian manufacturers, in spite of everything, can be proud that they have created the most "compact" supercomputer in the world!

Flops is a unit that represents the performance of a supercomputer. One petaflops (1 Pflops) means that the machine can perform 1 quadrillion (1,000 trillion) operations per second. Now only two machines have a capacity of more than 1 Pflops - the Jaguar, assembled by Cray, and the Roadrunner, manufactured by IBM. Both supercomputers are located in the United States. In general, out of the top ten, only two supercomputers are located outside the United States: in Germany and China.

04.08.2009 12:20

Today, the computer industry is at the forefront of science and technology. To solve complex problems in the field of physics, astronomy, biology, medicine, large computing power is required. It is supercomputers that can help with this, because they are created for this.

Recently, information has appeared quite often that another supercomputer has been created somewhere. But what is this miracle of technology? In the modern sense, a supercomputer is a powerful electronic computing machine with a performance of over one trillion floating point operations per second or teraflops. Flops (from the English. Floating point Operations Per Second) is a value for measuring the performance of computers, showing how many floating point operations per second a particular computing system performs. As a rule, a modern supercomputer is a multiprocessor or multicomputer complex (and in some cases a combined version) operating on a common memory and a common field of external devices.

Traditionally, the main field of application of supercomputers is scientific research. Plasma physics and statistical mechanics, condensed matter physics, molecular and atomic physics, the theory of elementary particles, gas dynamics and the theory of turbulence, astrophysics are just some of the areas where enormous computer power is involved.

Today, super-powerful computer systems are also used to solve technical problems. These are, first of all, the tasks of the aerospace and automotive industry, nuclear energy, prediction and development of mineral deposits, the oil and gas industry, as well as the construction of a supercomputer directly.

Supercomputers are also traditionally used for military purposes. In addition to developing a variety of weapons, they simulate their use. For example, in the United States, the computing power of the Department of Energy's supercomputer will be required to simulate the use of nuclear weapons, which will make it possible to completely abandon real nuclear tests in the future.

Currently, most of the TOP-500 supercomputers are engaged in scientific research. In this area, 72 most powerful information and computing machines are involved. The financial industry is served by 46 supercomputers, 43 machines serving geophysics, 33 working in information services, 31 managing logistics, 29 developing semiconductors, 20 producing software, 18 used in information services, and 12 systems running the Internet.

Working with huge arrays of computations distinguishes supercomputers from servers and mainframes - computer systems with high overall performance designed to solve typical tasks, for example, maintaining large databases or working with multiple users at the same time.

An increase in the performance of computing systems occurs primarily due to an increase in the speed of the physical and technological base ( electronic components, memory devices, means of communication, input-output and display of information) and the development of parallelism in the process of information processing at all system-structural levels, which is associated with an increase in the number of components involved (processing elements, memory volumes, external devices).

The most popular supercomputer architecture (72% in the TOP-500 list) today are the so-called clusters. To build a cluster architecture of a supercomputer, computing nodes are used, which are sometimes the most ordinary computers. In such a node, there are usually several processors - from 2 to 8. For this, quite ordinary components widely available on the market are used - motherboards (SMP-multiprocessor) boards, processors from Intel, AMD or IBM, as well as ordinary RAM modules and hard drives.
In their relatively short history, supercomputers have evolved from low-power systems by modern standards to machines with fantastic performance.

The very first mention of a supercomputer dates back to the end of the 20s of the last century, when this term appeared on the pages of the New York World newspaper in the form of the phrase "super computing" (translated from English - supercomputing). This concept referred to tabulators - electromechanical computers manufactured by IBM for and for the needs of Columbia University and producing the most complex calculations for that time. Naturally, then there were simply no supercomputers in the modern sense, this distant ancestor of modern computers was rather a kind of calculator.

The reference to the term "supercomputer" in relation to a powerful electronic computing system is attributed to George A. Michael and Sidney Fernbach, employees of the Livermore National Laboratory (USA, California) and Control Data Corporation. At the end of the 60s, they were engaged in the creation of powerful computers for the needs of the US Department of Defense and the energy industry. It was at Livermore Laboratory that most supercomputers were developed, including the fastest supercomputer in the world from 2004 to 2008, the Blue Gene / L.

However, the term "supercomputer" came into widespread use thanks to the American computer developer Seymour Cray, who back in 1957 created the Control Data Corporation, which was engaged in the design and construction of electronic computing systems that became the ancestors of modern supercomputers. In 1958, under his leadership, the world's first powerful computer based on CDC 1604 transistors was developed. It is worth noting that Seymour Kray's company was the first to mass-produce supercomputers - in 1965, the CDC-6600 machine with a capacity of 3 million operations per second entered the market. This computer became the basis for a whole direction that Cray founded in 1972 and called Cray Research. This company was exclusively engaged in the development and production of supercomputers. In 1976, Cray Research released the CRAY-1 computing system with a speed of about 100 megaflops. And nine years later, in 1985, the CRAY-2 supercomputer overcomes the computation speed of 2 gigaflops.

In 1989, Seymour Cray opens Cray Computer Corporation with a clear focus on supercomputer market prospects. Here he creates the CRAY-3 supercomputer, the speed of which has already reached five gigaflops. Associated with this computer interesting fact... The fact is that after the appearance of CRAY-3, the expression "Cray time" entered the English language, which meant the cost of an hour of operation of a supercomputer (at that time it was $ 1,000 per hour). There is another expression that went around in the circles of computer specialists - "A supercomputer is any computer that Seymour Cray created."

It is worth noting that in the 80s of the XX century, many small competing companies appeared that created high-performance computers. But by the mid-90s, unable to withstand the competition with large corporations, most small and medium-sized firms left this field of activity.

Today supercomputers are unique systems created by "traditional" computer players such as IBM, Hewlett-Packard, Intel, NEC and others. Exactly these computer giants are now dictating the rules of the game in high performance electronic computing systems.

In 1997, the American Intel released its ASCI Red supercomputer, which became the world's first system with a speed of more than one trillion operations per second - 1.334 teraflops. Intel supercomputers held the lead for another two years, but in 2000 the first was IBM's ASCI White computer installed at the Lawrence Livermore National Laboratory, which produced $ 4 trillion every second. 938 billion calculations (4.938 teraflops). This supercomputer held the leading position for another year, having received a speed equal to 7.226 teraflops after the upgrade. But already in April 2002, the Japanese company NEC announced the launch of the Earth Simulator supercomputer, which was able to reach a maximum speed of 35.86 teraflops.

The world of supercomputers went through another change of leaders in the fall of 2004 - on September 29, the supercomputer of the IBM Blue Gene / L company took the first place in the world. This powerful computing system reached a speed of 36.01 teraflops. However, this record did not last long - already on October 26, the US National Aeronautics and Space Administration (NASA) announced that its new supercomputer Columbia, built by Silicon Graphics and named after the shuttle that died in February 2003, performed a series of calculations at a speed 42.7 teraflops. A few days later, the same computer was able to increase the speed to 51.87 teraflops.
In early November 2004, the title of absolute record holder was again won by the Blue Gene / L, another sample of which was released by IBM for the US Department of Defense. Currently, the maximum speed of its work exceeds 70.72 teraflops. This computer was the leader until June 2008, when IBM for the nuclear laboratory in Los Alamos (USA, New Mexico) built its next supercomputer masterpiece - the most powerful roadrunner electronic computing system ever created.

Especially for accounting of supercomputers, the TOP-500 project was established, the main task of which is to compile a rating and descriptions of the most powerful computers in the world. This project was opened in 1993 and publishes an updated list of supercomputers twice a year (in June and November).

So, as already mentioned, the most powerful supercomputer today, according to the latest edition of the TOP-500 rating, is the IBM Roadrunner computing system. This computer is built on a hybrid scheme of 6500 dual-core AMD processors Opteron and nearly 13,000 IBM Cell 8i processors housed in dedicated TriBlades racks connected by Infiniband - a high-speed dial-up serial bus... Its peak performance is 1.105 petaflops.

Roadrunner runs Linux. The supercomputer from IBM occupies about 1,100 square meters of space and weighs 226 tons, and its power consumption is 3.9 megawatts. The cost of the IBM Roadrunner was $ 133 million.

The US Department of Energy will use RoadRunner to calculate aging of nuclear materials and analyze the safety and reliability of a nuclear arsenal. In addition, this supercomputer will be used for scientific, financial, transportation and aerospace computing.
The second place in the ranking is occupied by the Cray XT5 Jaguar supercomputer, which is installed in the laboratory of the US Department of Energy in Oak Ridge, Tennessee. Its performance is 1.059 petaflops.

Jaguar set a new performance record after adding two hundred Cray XT5s to its 84 Cray XT4s. The latter are based on AMD Opteron quad-core processors. Each Cray XT5 block contains up to 192 processors. The total number of Jaguar processors is 45 thousand.

Of the others technical characteristics of the supercomputer, the amount of its RAM and the capacity of disk drives are known, they are equal to 362 terabytes and 10 petabytes, respectively.

Unlike the IBM Roadrunner, Jaguar's supercomputer has to tackle peaceful tasks. For example, it will be used to simulate climate change in areas such as renewable energies and materials science. In addition, the US Department of Energy says that Jaguar will allow research into processes that were previously out of the question. What these processes, unfortunately, is not reported.

The third most powerful supercomputer in the world, as well as the fastest in Europe, is the IBM Blue Gene / P model of the supercomputer line, which is installed in the research center of the city of Julich in Germany. Computing complex JUGENE, which was launched this summer, has 72 racks, which house 294,912 PowerPC 450 core 850 MHz processors, and its power is 825.5 teraflops. The memory capacity of the German supercomputer is 144 TB. In addition, this supercomputer is one of the most economical devices among similar solutions - its power consumption is about 2.2 MW.

The computational resources of this supercomputer are used, among other things, in the calculation of projects related to thermonuclear research, the development of new materials, the search for next-generation drugs, as well as in modeling climate change, the behavior of elementary particles, complex chemical reactions, etc. the projects are handled by a group of independent experts.

By the way, as of November 2008, Russia ranks 11-14 in terms of the number of installed systems, along with Austria, New Zealand and Spain. The USA is the leader in this indicator, where there are about 300 supercomputers from the rating. However, in terms of power, the most productive Russian supercomputer MVS-100K, which performs tasks in the Interdepartmental Supercomputer Center of the Academy of Sciences of the Russian Federation, is only in 54th place. Despite this fact, MVS-100K with a peak performance of 95.04 teraflops per currently is the most powerful supercomputer installed in the CIS countries. It consists of 990 compute modules, each of which is equipped with two Intel Xeon quad-core processors, clocked at 3 GHz. In the near future, it is planned to increase the performance of MVS-100K up to 150 TFlops. This supercomputer is designed to solve a wide range of complex scientific and technical problems.

What are the prospects for supercomputers in the future? According to experts, the most rosy ones. But it is already clear that their performance will grow rather quickly due to the increase in the number of processor cores and the average frequency of processors. In addition, to solve applied problems in supercomputers, not only general-purpose processors will be used, but also specialized ones (for example, graphics processors developed by Nvidia and ATI) designed for specific tasks... Also, supercomputer manufacturers will look for new unique architectural solutions that would not only increase the power of computers, but would also give advantages in competition in the commercial market. In addition, in the future, supercomputers will noticeably increase the efficiency due to the development of software tools. The intellectual abilities of supercomputers will also increase, and along with this, the professional qualities of programmers and other IT specialists will grow.

It is also worth noting that in the future, high-performance computing systems will gradually increase their presence in the global computer market. According to IDC, the global supercomputer market is growing at an annual rate of 9.2%. Revenue of supercomputer manufacturers in the second quarter of 2008 amounted to 2.5 billion dollars, which is 4% more than the same period last year and 10% more than in the first quarter of 2008.

According to IDC analysts, HP took the first place in terms of revenue with a market share of 37%, followed by IBM (27%) and closes the "three" leaders Dell (16%). According to the forecast of analysts IDC, the market for supercomputers by 2012 will reach 15.6 billion dollars.

Of the systems presented in the TOP-500, 209 (41.8%) were manufactured by HP specialists. IBM is in second place with 186 computers, while Cray is in third with 22 supercomputers.

As for Russia, according to Mikhail Kozhevnikov, commercial director of the T-Platforms company, the annual growth in the supercomputer market is about 40%. Thus, according to T-Platforms, the volume of the supercomputer market in Russia in 2007 amounted to about $ 60 million, and in 2008 the market grew to about $ 80 million. According to Mikhail Kozhevnikov, even during the crisis, it is expected that in 2009 the market will grow by about 60%, and under favorable conditions, even up to 100%.

As you can see, supercomputers are only gaining "commercial" momentum. It is difficult to imagine, but, indeed, bulky computers are sold like "hot cakes" in the computer market. Should we expect a smaller version of a supercomputer with the same high performance that large computing systems now have? Probably, only supercomputers themselves can answer this difficult question, because this is their job.

End of work -

This topic belongs to the section:

New list of the most powerful supercomputers named

The creation of super-powerful computers was named one of the priorities of the technological .. authors Pavel Lebedev the creation of super-powerful computers was named one of the priorities ..

If you need additional material on this topic, or you did not find what you were looking for, we recommend using the search in our base of works:

What will we do with the received material:

If this material turned out to be useful for you, you can save it to your page on social networks:

− 10 21 yottaflops − 10 24 xeraflops − 10 27

FLOPS(or flops or flop / s) (acronym from English. Fl oating point O perations P er S econd , pronounced like flops) is a value used to measure the performance of computers, showing how many floating point operations per second a given computing system performs.

Insofar as modern computers have a high level of performance, derivatives of FLOPS are more common, formed by using standard SI prefixes.

Flops as a measure of performance Like most other performance indicators, this value is determined by running a test program on a test computer that solves a problem with a known number of operations and calculates the time it took to solve it. The most popular performance test today is the LINPACK program, which is used, among other things, to compile the TOP500 supercomputer rating. One of the most important advantages of the flops indicator is that, up to some limits, it can be interpreted as an absolute value and calculated theoretically, while most other popular measures are relative and allow you to evaluate the system under test only in comparison with a number of others. This feature makes it possible to use the results of the work of various algorithms for evaluation, as well as evaluate the performance of computing systems that do not yet exist or are under development. Applicability limits Despite the seeming unambiguity, in reality flops is a rather bad measure of performance, since its very definition is ambiguous. Under the "floating point operation" can hide a lot of different concepts, not to mention the fact that the significant role in these calculations is played by the width of the operands, which is also not specified anywhere. In addition, the amount of flops is influenced by many factors that are not directly related to the performance of the computing module, such as: the bandwidth of communication channels with the processor environment, the performance of the main memory and the synchronization of the cache memory at different levels. All this ultimately leads to the fact that the results obtained on the same computer using different programs may differ significantly, moreover, with each new test, different results can be obtained using the same algorithm. This problem is partly solved by the agreement on the use of monotonous test programs (the same LINPACK) with averaging the results, but over time the capabilities of computers "outgrow" the scope of the accepted test and it begins to give artificially low results, since it does not use the latest capabilities of computing devices. And for some systems, generally accepted tests cannot be applied at all, as a result of which the question of their performance remains open. Reasons for widespread use Despite the large number of significant shortcomings, the flops indicator continues to be successfully used to assess performance based on the results of the LINPACK test. The reasons for this popularity are due, firstly, to the fact that flops, as mentioned above, is an absolute value. And, secondly, many problems of engineering and scientific practice, ultimately, are reduced to solving systems of linear algebraic equations, and the LINPACK test is precisely based on measuring the speed of solving such systems. In addition, the vast majority of computers (including supercomputers) are built according to the classical architecture using standard processors, which allows the use of generally accepted tests with high reliability. As shown on processors Intel Core 2 Quad Q9450 2.66GHz @ 3.5GHz and Intel Core 2 Duo E8400 3000 MHz (2008) LINPACK does not use algebraic expression solutions, since any operation cannot go faster than 1 processor clock. So for Intel Core 2 Quad processors, one clock requires one or two hertz. Since for floating point tasks: division / multiplication, addition / subtraction - much more than one clock cycle is required, it can be seen that these processors could not produce 48 Gigaflops and 18.5 Gigaflops, respectively. Often, instead of a floating point division operation, loading data in DMA mode from RAM into the processor stack is used. This is how LINPACK works in some tests, but strictly speaking, the result is not a flops value. Note: the remark about the impossibility of performing more than one operation per cycle is absolutely incorrect, since all modern processors in each of their cores contain several execution units of each type (including those for floating point operations) working in parallel and can execute more than one instruction per cycle. This feature architecture is called superscalar and first appeared in the very first processor Real System Performance Review Due to the high scatter of LINPACK test results, approximate values ​​are given, obtained by averaging indicators based on information from different sources. The performance of game consoles and distributed systems (with a narrow specialization and do not support the LINPACK test) are given for reference purposes in accordance with the numbers stated by their developers. More accurate results indicating the parameters of specific systems can be obtained, for example, on the website. Supercomputers Personal computers Processors Intel Core 2 Duo E8400 3.0GHz () - 18.6 Gflops Using LINPACK 10 standard version Intel Core 2 Duo E8400 3.0GHz @ 4.0GHz () - 25 Gflops (LINPACK Benchmark 10.0 64-bit) in Windows Vista x64 Ultimate SP1 Intel Core 2 Quad Q9450 2.66GHz @ 3.5GHz - 48GFlops (LINPACK Benchmark 10.0 64-bit) on Windows 2003sp2 x64 Pocket Computers Distributed systems Gaming consoles Man and calculator Notes (edit) see also Links TOP500 Supercomputer rating TOP500 (eng.) The Performance Database Server Large Computing Performance Database Roy Longbottom's PC Benchmark Collection A selection of PC benchmarks (including LINPACK) and test results (eng.) Linpack CPU Benchmark for Pocket PC LINPACK version for Pocket PC Wikimedia Foundation. 2010. See what "Petaflops" is in other dictionaries: Client screenshot [email protected] for PlayStation 3, showing a 3D model of a simulated protein Type Distributed Computing ... Wikipedia

Sony Computer Entertainment Inc. proudly announced that the participation of the PLAYSTATION 3 entertainment system enabled the project [email protected] Stanford University to achieve aggregate power in excess of 1 petaflops.

Petaflops is the ability of a computer or network to perform 1 quadrillion (one followed by 24 zeros) floating point calculations per second (FLOPS). In other words, if every person on Earth carried out a simple mathematical calculation (for example, calculating a percentage of a certain amount), then every earthling would need to do 75,000 simple mathematical calculations per second in order for the total computing power of humanity to reach petaflops.

A similar increase in the computing power of the project [email protected] will significantly accelerate research that previously took decades. All of this is made possible by the PLAYSTATION 3's Cell Broadband Engine (Cell / B.E.), Which has over 180 GFLOPS (billions of floating point operations per second) processing power. Cell / B.E. about 10 times faster than a conventional PC processor, so PLAYSTATION 3 can literally be called a home supercomputer. PLAYSTATION 3's participation in the project helps scientists identify the causes of diseases such as Parkinson's disease, Alzheimer's disease and cancer.

According to Stanford University Adjunct Professor of Chemistry and Project Leader [email protected] Vijay Pande, adding PLAYSTATION 3 to the project [email protected] put at the disposal of scientists such a power that they could not even dream of.

In turn, the president and CEO of the American division of SCEI Jack Tretton said that even at the development stage, the company's engineers knew that the power of PLAYSTATION 3 would be used not only for entertainment, but also for the benefit of all mankind. For the entire SCEI team, the use of her brainchild in projects like [email protected], - this is a reason for pride.

Protein research is an extremely complex process. Have ordinary computer solution the simplest task can take up to 30 years. [email protected] distributes computations among thousands of computers connected in a single network. Until recently in [email protected] used only personal computers... The project involved about 200 thousand PCs, the total capacity of which was about a quarter of a petaflops. Thanks to a firmware update on March 15, 2007, PLAYSTATION 3 "learned" how to work with the project. After that in [email protected] more than 600 thousand PLAYSTATION 3 users have registered, which allowed them to exceed the power mark of 1 petaflops.

To take part in [email protected], you just need to connect PLAYSTATION 3 to the Internet, download the new version of the internal software System Software and click the icon [email protected] For more information, see the Network section of the XMB Main Menu (XrossMediaBar). In the settings, you can set the option automatic start annexes [email protected] while PLAYSTATION 3 is in standby mode. For the application to launch automatically, the PLAYSTATION 3 must be turned on and connected to the Internet.

It should be noted that [email protected] Is just the beginning. SCEI plans to add support for many other distributed computing projects to PLAYSTATION 3 in a variety of scientific fields, from medicine to social and environmental research. At the same time, PLAYSTATION 3 owners will be able to determine for themselves where to use the power of their entertainment system.