Sitemap
Arjun Thakur

Arjun Thakur

Supercomputers

13 min readMay 1, 2022

--

Authors: Shubham Shah, Shruti Singh, Akash Sinha, Vishal Sirvi, Shreyansh Srivastava and Arjun Thakur.

What makes a Computer “Super”?

Press enter or click to view image in full size

Faster processing speeds, extra memory, and super-sized storage capacity are what make supercomputers the tools of choice for many researchers. The devices we keep at home and in our pockets are pretty advanced compared to the technology of the past, but they can’t hold a candle to the raw power of a supercomputer.

A supercomputer is a computer with a high level of performance as compared to a general-purpose computer. The performance of a supercomputer is commonly measured in floating-point operations per second (FLOPS) instead of million instructions per second (MIPS). It performs at or near the highest operational rate for computers.

Traditionally, supercomputers have been used for scientific and engineering applications that must handle massive databases, do a great amount of computation or both. Advances like multicore processors and general-purpose graphics processing units have enabled powerful machines that could be called desktop supercomputers or GPU supercomputers. Supercomputers have very high storage capacity and can store thousands of times more data than any average personal computer. The supercomputer can process billions of instructions in a second. Modern supercomputers have thousands of processors and are used for complex computations.

Processing power is the main difference between supercomputers and general-purpose computer systems. A supercomputer can perform 100 PFLOPS. A typical general-purpose computer can only perform hundreds of gigaflops to tens of teraflops. Supercomputers consume lots of power. As a result, they generate so much heat that they need to be stored in cooling systems.Both supercomputers and general-purpose computers differ from quantum computers, which operate based on the principles of quantum physics.

The 3 fastest supercomputers in the world

Fugaku (Japan)

Jointly developed by RIKEN and Fujitsu, Japan’s Fugaku is the new number one fastest supercomputer in the world. Japan has not had a system take the top spot since June 2011 when Fugaku’s predecessor, the K computer, debuted in first place. Boasting nearly 7.3 million cores and a speed of 415.5 petaFLOPS, Fugaku far outperforms previous #1 Summit’s 148.6 PetaFLOPS, bringing HPC technology one step closer to the promised exascale era.

Press enter or click to view image in full size
FUGAKU SUPERMCOMPUTER

Summit (US)

Poised at the Summit. The US Department of Energy’s Oak Ridge National Laboratory (ORNL) in Tennessee slips to #2 after two years in first place. Courtesy ORNL.As part of the US Department of Energy’s renewed commitment to supercomputing power, Oak Ridge National Laboratory’s (ORNL) Summit first claimed the #1 spot in June 2018, taking the top rank from China for the first time in 6 years.

Press enter or click to view image in full size
SUMMIT SUPERCOMPUTER

Sierra (US)

Lawrence Livermore National Laboratory’s (LLNL) Sierra initially debuted at #3 on the June 2018 list with 71.6 petaFLOPS. Optimization later pushed the processing speed on its 1,572,480 cores to 94.6 petaFLOPS, earning it the #2 spot in November 2018. However, the ascension of a new number one in June 2020 pushes Sierra back to third position.

Press enter or click to view image in full size
SIERRA SUPERCOMPUTER

Evolution of Supercomputers

The United States, China, the European Union and Japan are already competing to develop supercomputers that are faster, more powerful, and technologically advanced. However, the US marked a big stride in the history of supercomputers by introducing CDC (Control Data Corporation) 6600, in 1964.

Early Supercomputers

From the mid-1980s through the early 1990s, Cray dominated supercomputers. The Cray Y-MP, with up to eight 32-bit vector processors operating at 167MHz, was introduced by Cray Research in 1988. It was the first supercomputer to sustain more than 1GFLOPS (109 floating point operations per second) and provided SRAM main memory options of 128, 256, or 512MB.

Press enter or click to view image in full size
Cray Y-MP Supercomputer

The X-MP, a forerunner of the Y-MP, was marketed to a West German nuclear research organization worth $11.4 million in 1981, or $32.6 million in 2020 dollars. Cray may have dominated the supercomputing business in the early 1990s, but they were far from alone. NEC’s SX line of vector supercomputers was named after it. The first two NEC models, the SX-1 and SX-2, were introduced in 1985. Both systems have up to 256MB of main memory. The SX-2 was the first supercomputer to surpass 1 GFLOPS, according to claims. It has four high-performance vector operation pipelines, each with up to 16 arithmetic units capable of performing multiple/parallel operations. The NEC SX-1 cost about half as much as the SX-2 and offered about half the performance.

Supercomputers in the Latter 1990s era

NEC introduced the SX-5 system in 1998. It has the ability to attain a performance of 4 TFLOPS. Each system node contains up to 128GB of primary memory and 16 CPUs. Up to 32 nodes might be linked together for a total memory capacity of 4,096GB.

Press enter or click to view image in full size
NEC SX Vector Supercomputer

SGI introduced the SGI Origin 2000, a large-scale HPC system, in 1996. It could handle CPUs with a performance range of 2 to 128 MIPS. These were MIPS R10000 processors, which started at 180MHz and increased to 300 and 400MHz over time. SGI had the advantage of having non-supercomputer CPUs, such as the Origin 2000. They were also used in SGI workstations. Both “traditional” supercomputers and PC systems had a key decade in the 1990s. Both had made rapid progress, setting a course for the early 2000s.

Evolution of Indian Supercomputer

Dr. Vijay Bhatkar is known as the “Father of Indian Supercomputers’ ‘, he introduced the Indian PARAM Supercomputer. In November 1987, the Centre for Development of Advanced Computing (C-DAC) obtained a three-year budget of Rs 375 million to create supercomputers with 1000Mflops (1 Gflop). Three C-DAC flights revealed the “PARAM’’ (Parallel Machine) supercomputer family.

Press enter or click to view image in full size
Development of PARAM supercomputer over the years

PARAM 8000

In August 1991, the PARAM 8000, a 64-node system, was introduced as the first machine designed from the ground up. This 256-node system has a theoretical performance of 1 Gflops, but in actuality, it only has roughly 100–200Mflops. The PARAM 8000 was built on a distributed memory MIMD architecture with a reprogrammable connectivity network.

PARAM 8600

PARAM 8600 was released in 1992 as an upgrade to PARAM 8000. The Intel i860 CPU was chosen by C-DAC as a way to increase power. Each 8600 cluster was equivalent to four PARAM 8000 clusters in terms of processing power.

PARAM 9000

The PARAM 9000 was initially shown in 1994, and it was created to combine cluster and massively parallel processing computer tasks. Employing the Clos network design, this system expanded up to 200 CPUs using 32–40 processors. The SuperSPARC II processor was used in the PARAM 9000/SS; the UltraSPARC processor was used in the PARAM 9000/US; and the DEC Alpha was used in the PARAM 9000/AA.

PARAM 10000

In 1998, the PARAM 10000 was revealed. This supercomputer featured many independent nodes, each of which was based on a Sun Enterprise 250 server with two 400 Mhz UltraSPARC II processors. This system’s top speed was 6.4 Gflops. This would have 160 CPUs and 100 Gflops of performance, readily expandable to the Tflop range.

Get Arjun Thakur’s stories in your inbox

Join Medium for free to get updates from this writer.

Later on, PARAM Padma, PARAM Yuva, PARAM Brahma, PARAM Ishan and PARAM Siddhi-AI were introduced which were known to be technologically advanced successors of each other and have high computing power. The PARAM Siddhi-AI supercomputer is India’s fastest, with a peak Rpeak of 5.267 Pflops and a maximum Rmax of 4.6 Pflops (Sustained). The AI assists research in innovative materials, combinatorial chemistry, and astrophysics, as well as universal healthcare, disaster prevention, and the Covid-19 application, by allowing for faster simulations, diagnostic imaging, and genome sequencing. In November 2020, PARAM Siddhi-AI was ranked 63rd amongst world’s best supercomputers. NVIDIA DGX SuperPOD networking is used, as well as C-own DAC’s HPC-AI engine, software frameworks, and cloud platform.

Different types of Supercomputers

There are two types of supercomputers which evolve gradually

  • General Purpose Supercomputers.
  • Special Purpose Supercomputers.
SUPERCOMPUTER

1. General-purpose supercomputer

Further General Purpose supercomputers can be classified as:

  • Vector processing computers: Vectors or array processors are used in a vector processing computer. These processors are similar to CPUs that can assist in performing mathematical operations on a large number of data pieces.
  • Commodity computers: Computers that are a collection of commodity computers or conventional PCs connected to a local area network with low latency and high bandwidth.
  • Tightly connected cluster computers: A cluster supercomputer is a collection of closely linked computers that operate as a single entity.

2. Special purpose supercomputers

Supercomputers that are developed specifically to accomplish a specific task or aim are known as special purpose computers. They frequently employ Application-Specific Integrated Circuits (ASICs), which outperform general-purpose supercomputers. These supercomputers are used for brute-force codebreaking and astrophysical computing while their hardware architecture is designed to solve a single task.

Supercomputers and Artificial Intelligence

Supercomputers have the potential to boost the speed of artificial intelligence systems.

However, because super-computing is commonly used for data analysis and scientific activities like processing massive amounts of data to solve clinical, environmental, infrastructure, and a variety of other scientific problems, few members of the general public have a thorough understanding of how advanced technology affects their lives.

Supercomputers have incredible processing speed, allowing them to convert simple data into useful data in seconds, minutes, or days, rather than the years or decades it would take if done manually.

The world’s new fastest supercomputer, according to Data Center Knowledge, is called Fugaku and is powered by Fujitsu’s 48-core A64FX SoC. The RIKEN Center for Computational Science (R-CCS) in Kobe, Japan, has it installed. The non-profit Top500 organization announced on June 22, 2020 that it outperforms the previous performance champion, the US Department of Energy’s Summit system, which topped the Top500 list twice in a row.

Artificial intelligence (AI) is putting machine-assisted capabilities to the test. It has the potential to boost the speed with which machines perform human-like tasks. Because of its automation and advanced analytics, AI is becoming more important. Using the power of machine learning, deep learning, and natural language processing, AI is providing tremendous benefits to organizations. Artificial intelligence (AI) helps companies capitalize on new digital industry trends. Artificial intelligence will benefit individuals, markets, and society as a whole.

Meta’s AI Supercomputer

Although supercomputers have long been required in fields like physics and space science, the increased use of artificial intelligence and machine learning has resulted in a surge in demand for supercomputers capable of performing quadrillions of calculations per second. In reality, exascale supercomputers, the next generation of supercomputers, are improving efficiency in these areas.

Supercomputers, or, to put it another way, machines with accelerated hardware, are deserving of being used to increase the speed of artificial intelligence systems. Because of its improved pace and ability, it can train faster on larger, more detailed sets, as well as more oriented and deeper training sets.

According to AIM, Microsoft announced last year that it had developed a supercomputer in collaboration with OpenAI and would be hosted in Azure Cloud. The supercomputer, according to the company, has over 2,85,000 CPU cores, 10,000 GPUs, and 400Gbps network connectivity for each GPU server, allowing it to train various artificial intelligence models.

Supercomputing Challenges

Most of the challenges and issues related to supercomputing are because of the hardware and software interrelationships and their algorithms. However there are some other challenges as well such as movement of data , fault tolerance, power consumption, and extreme parallelism. In this section we will be talking about all these challenges and issues in brief.

Speed of data moment

This challenge basically refers to the lack of available technology which are mainly used to fetch and retrieve data at sufficiently high rates. For faster computation, fast execution speed and data transfer rate is very important and that too has to be compatible with the given power packet which gives rise to the other challenge about which we will discuss further. Actually in most of the cases it has been found that the time and energy required for data movement from main memory to the processing unit and from processing unit to main memory again is greater in magnitude than the energy and time required for the actual computation and operation.

Fault Tolerance

Fault tolerance is the ability of a system to continue the given computational task without being affected by any available faults in the program. Incase of supercomputing where a huge number of components and peripherals are being interfaced with larger systems along with advanced technologies being operated at specific and limited voltage levels, providing this feature is yet another challenge.

Power Consumption

As discussed earlier, power consumption rates along with the speed of data transfer are the most basic and yet challenging issues faced in the field of supercomputing. As per findings, average power consumption of an exaFLOP built with the existing available technologies would be more than 600 megaWatt. This will not only require a more powerful source of energy but also increase the cost of computing.

Level of Parallelism

For faster execution, greater no of threads along with multilayer parallelism is needed. Infact, if the computation rate is to be as fast as 1 exaFLOP, the requirement is of 1 billion floating point units, where each unit is performing 1 billion calculations per second. As far as parallelism is concerned, the requirement is parallel processing of billion threads being solving a small segment of a single problem.

Future of Supercomputers

As we all know, the first supercomputer appeared in the 1960s, principally developed by Control Data Corporation. This business led the market until its co-founder Seymour Cray departed to create his own company, Cray Research, in the 1970s. He subsequently dominated the market with his innovative ideas, dominating supercomputing for several years. A number of smaller rivals joined the market in the 1980s, but many vanished after the mid-1990s supercomputer market meltdown.

After a quick overview of what supercomputers were and when they first appeared, let us attempt to predict their future. For this, we will need to refer to a supercomputing performance metric known as floating point operations per second (FLOPS). FLOPS, or floating-point operations per second, is the unit of measurement used to quantify a supercomputer’s performance capabilities.

Sharing some FLOPS statistics from previous generations of supercomputers. For example, Intel’s ASCI Red supercomputer was the first to reach one teraFLOPS (a trillion computations per second) in 1997. It had a peak performance roughly 800 times faster than the Cray-2, which had been introduced only 12 years before in 1985. In 2008, IBM’s Roadrunner reached one petaFLOPS (1,000 trillion computations per second), which is around 1,000 times faster than ASCI Red.

Supercomputers have continued to advance by orders of magnitude since then, but at a slightly slower rate especially from the second half of 2013, according to the TOP500 list kept by academics in the United States (which employs the LINPACK benchmark) and on which the graph stated below is based on. Nonetheless, the most powerful machines will soon surpass the next milestone: exaFLOPS. In other words, 1,000,000,000,000,000,000,000,000 floating point operations per second. Several candidates are in the wings, and one of them may win the championship in the near future.

Supercomputer Future Prediction

Some future supercomputers are already in the works. For example, Oak Ridge National Laboratory in Tennessee is creating a $600 million machine named Frontier, which is expected to arrive by the end of 2021 and has a goal capability of 1.5 exaFLOPS. Meanwhile, on the suburbs of Chicago, the Argonne National Laboratory is creating Aurora, a system created by Intel and Cray that will have more than 200 cabinets with a combined 1 exaFLOPS. This was supposed to be the world’s first exaFLOPS computer, but owing to chip production delays with Intel, it will now arrive after Frontier.

Isn’t the race to become a worldwide supercomputing power fierce? So, let’s see who is currently at the top of the world rankings. China now leads the TOP500 list, with 188 computers (compared to 122 for the United States), although it only has two machines in the top ten. Japan already boasts the world’s fastest supercomputer, Fugaku, which has a speed of 442 petaFLOPS. As a result, the first exaFLOPS machine will require at least 2.3 times that speed.

Press enter or click to view image in full size
FUGAKU SUPERCOMPUTER

Looking even farther forward, we reach yottaFLOPS, a computing milestone that might occur in the early 2050s. A supercomputer with yottaFLOPS would be 1,000 times faster than a zettaFLOPS machine and a million times faster than an exaFLOPS machine. To put this in context, a yottaFLOPS supercomputer could finish a large-scale simulation that would have taken six months to perform in 2021 in only 15 seconds.

What can such speed and computational power get us? The uses of such a machine might open up a whole new world of possibilities, which are part of the ‘Singularity’ predicted by many futurists. We can only speculate on the kinds of game-changing discoveries that will result from this technology and its exponentially increased possibilities. Perhaps we will be able to reconstruct the early moments of the Big Bang in minute detail. Or perhaps a new material will be discovered at the nanoscale that is far more sophisticated than anything now available.

Along with the development of the first super AI, with a supercomputer as its brain, is an even more fascinating (and terrifying) potential. According to one scenario foreseen by futurists such as Life 3.0 author Max Tegmark, this organism may attempt to escape its confinement and spread over the web. It might then download itself onto robot bodies, resembling the premise of a science fiction film. It remains to be seen if this will be beneficial or bad for humanity.

Give a couple of claps if you found the article useful. Thank You and Happy Learning !!!

--

--