What is the Top500 list of supercomputers

Surely you have heard something about the Top500 list . Well, in this article we are going to tell you everything you need to know about this famous list that is responsible for classifying the 500 most powerful supercomputers (HPC) in the world every year.

Index of contents

• What is the Top500?
  • A little history
  • Interesting facts from the Top500
  • Example Current Top10
• What is Green500?
• What is Graph500?

What is the Top500?

This project arises to classify the 500 most powerful supercomputers in the world, hence its name Top500 . The first time it was done was in 1993, and since then it has published an updated list twice a year. The first is held in June and the second in November. These are not dates chosen by chance, since June is presented at the International Supercomputing Conference and November is at the ACM/IEEE Supercomputing Conference.

Anyone interested in knowing these lists and all the details can access them for free from their official portal top500.org . Also, you should know that this project is not only for countries to get muscle and show off their computing power, it is also a way to track and follow up on HPC trends.

As for the confession of the lists, they are carried out by a group of experts and are based on the HPL classifications, that is, a portable implementation of the high-performance LINPACK benchmark. This benchmark is written in the Fortran programming language and can be run on distributed memory systems.

The 60th list, the last to be updated at the time of writing this article, was made in November 2022. Until June 2023 there will not be another updated one.

Although the list is made in many countries around the world , the undisputed leader in installed supercomputers was the United States, which has almost 50% of the total 500 machines analyzed at the moment. In second position was China, which has currently become the leader, being the first country in history to surpass the US, while other countries share the rest. Within Europe, Germany is the best positioned country, in third place, with more than 30 supercomputers, closely followed by Japan. And if you are wondering about Spain, the most powerful of the ones we have here and the only one that enters the Top500 list is the Marenostrum installed at the BSC in Barcelona.

However, in Spain there is a RES (Red Española de Supercomputación) , which is made up of several supercomputers distributed throughout the Spanish geography and interconnected with each other:

• Barcelona Supercomputing Center (BSC) – National Supercomputing Center (CNS) in Barcelona. Being the most powerful, it acts as a coordinator for the others.
• BIFI of the University of Zaragoza.
• PICASSO from the University of Malaga.
• NASERTIC of Navarre.
• CETA of the CIEMAT (Center for Energy, Environmental and Technological Research) in Madrid.
• Altamira of the University of Cantabria.
• Tirant from the University of Valencia.
• Autonomous University of Madrid.
• Institute of Astrophysics of the Canary Islands.
• FinisTerrae of the Supercomputing Center of Galicia.
• Castilla y León Supercomputing Center.
• Supercomputing Center of Extremadura.
• Consortium of University Services of Catalonia.
• Port of Scientific Information PIC.

Finally, it should be said that among the current experts involved in the compilation of this Top500 list we have Jack Dongarra from the University of Tennessee, Erich Strohmaier and Horst Simon from NERSC (National Energy Research Scientific Computing Center) and Lawrence Berkeley National Laboratory ( LBNL), and until 2014 Hans Meuer from the University of Mannheim in Germany also participated.

I also invite you to read our guide with the best processors on the market .

Some wonder if this list is reliable, and it really lists the 500 most powerful teams in the world. And the truth is that it cannot be assured with certainty. It could be said that it has the 500 most powerful supercomputers known, since if any country has a secret supercomputer, it would not appear on the list. Furthermore, it is known that there are some large machines that are not on this Top500 list, such as NSCA Blue Waters, OceanLight, RIKEN MDGRAPE, etc., that have not wanted to participate in the list…

A little history

Due to the rapid growth of supercomputing, it became necessary to track the capabilities of this sector for informational purposes. Also, in the early 1990s, a new definition of what a supercomputer was was needed. And after using various metrics that didn’t turn out to be appropriate until 1992, the idea of ​​using the installed processor count came up at the University of Mannheim .

In early 1993 Jack Dongarra was convinced to join the LINPACK benchmarks project and the first test for the list was produced. Consulting various sources, the first two lists were published and thus the Top500 site emerged as we know it today.

Since this date, number 1 in the Top500 ranking has always been occupied by a system that followed Moore’s famous law, doubling the number of transistors every 14 months or so. And, although most of these supercomputers were based on x86 processors such as the Intel Xeon and AMD Opteron/EPYC, as well as the IBM POWER, the truth is that lately we have also seen surprises with other more exotic CPUs, such as the Japanese Fugaku, with ARM-based Fujitsu A64FX chips.

In addition, systems based on heterogeneous computing have recently been added , so that the computing power of other units used as accelerators is also used, such as GPUs.

Interesting facts from the Top500

To better appreciate the evolution that the microprocessors used in HPC have had over the years in this Top500 list, you can see this graph:

As can be seen, in the current supercomputers on the Top500 list, Intel is the undisputed leader, followed by AMD and IBM POWER. Little by little, these have been strangling the rest of the diversity of processors that existed in the past, where everything was more distributed. However, little by little we are also seeing the appearance of Fujitsu A64FX, Cavium and ThunderX chips based on ARM, among other architectures.

On the other hand, the same analysis can also be done with operating systems, and in this case there is a clear dominant: Linux . And it is that the open source operating system is currently powering 100% of the supercomputers on the Top500 list:

Little by little, other Unixes have been giving way to Linux, which has killed them all. Within Unix there have been various operating systems, such as HP-UX, IBM AIX, Solaris, etc. And within Linux there are also various distributions such as:

Other curious facts is to know that the accelerators most used in these supercomputers on the Top500 list are:

As you can see, NVIDIA is the absolute leader in this type of systems for HPC.

On the other hand, if we focus on the manufacturers or suppliers of supercomputers , we find that the leaders are:

As we can see, the manufacturer Lenovo is one of the main ones . However, although they are not included in this table, there are also others such as IBM, Dell EMC, Fujitsu, NEC, MEGWARE, etc.

Example Current Top10

For example, to give you an idea, here is a table with the Top10 of the current Top500:

If you look at the previous table, we have several columns that I explain below :

• Position: It is the position you currently occupy within the Top500 ranking of the list as of the November 2022 revision.
• Rmax/Rpeak: refer to the LINPACK score, that is, the performance measured in PFLOPS or PetaFLOPS. While Rmax is the performance that has been achieved on LINPACK, Rpeak is the theoretical maximum performance that this system would have.
• Name: is the name of the supercomputer.
• Model: refers to the platform used for its creation or who sells it.
• CPU Core Count/Accelerator Core Count: The number of CPU or other accelerator processing cores it uses.
• Network interconnection: refers to the type of network that interconnects the installed nodes. They are usually high performance networks such as InfiniBand, Gigabit Ethernet or other proprietary ones. In this way, all these nodes can work as a single computer.
• Manufacturer: is the company in charge of its design and assembly.
• Site and Country: the place where the supercomputer is installed.
• Year: is the year of its installation.
• Operating system: the OS or SSOO that it uses to function.

What is Green500?

As you well know, these huge machines consume a lot of energy . These data processing centers could reach several tens of megawatt hours, which means a very high consumption and an equally expensive energy expense. And to all this we must add the energy necessary for the refrigeration systems, which is also usually several megawatts.

For this reason, every six months, the members of the Top500 list also show a list with the 500 supercomputers ordered according to their performance/watt consumed , that is, by their level of efficiency. This gives you an idea of ​​which are the most efficient systems.

The Green500 list has been topped by different types of teams, some based on x86, others on POWER, with or without accelerators, etc. However, in this case the ranking would be something similar to this example of the 3 most efficient supercomputers according to the Green500 list of 2022:

In this other case we have a column for the position occupied in the Green500, a second one that shows the efficiency accounted for in GFLOPS/w, then there is another column with the name of the supercomputer, another one with the technical characteristics, the manufacturer, the place where is installed and date, and performance based on Rmax in PFLOPS, to give you an idea of ​​where it would rank on the Top500 list.

As can be seen, the current number 1 of the Green500 can develop a performance per watt of 65,091 GFLOPS . On the other hand, if we look for it in the Top500 list, the Henri will be in the 11th position of the most powerful in the world.

You may be thinking, what about ARMs? Well, it seems that this is the world turned upside down, since while Arm has always boasted of performance and efficiency, and despite the fact that many saw it as crazy that an Arm could outperform other processors, the truth is that they have shown that those who believed that they were wrong. Case in point is the Fugaku, which is based on Fujitsu’s ARM chips and now sits at number 2 in the Top500, although not long ago it rose to number 1. However, when we look at the Green500 list, you’d think it agrees. the first…

Actually the Fugaku is in position 43, with 15,418 GFLOPS/w . So? Have you been cheated? Are ARMs not that efficient? Well, if we compare it with the Frontier that currently ranks 1 in the Top500, we see that it ranks much better in the Green500 ranking, with a sixth position. And that is based on AMD EPYC and AMD Instinct MI250X accelerators. And not only that, it has more cores, with a total of 8,730,112 compared to 7,630,848 for the Fugaku.

As can be seen in this comparison, ARM is not always the most efficient . Here’s the proof. Also, that thanks to the accelerators you can get great advantages in performance per watt consumed as I explained here . And, as can be seen in the Top500 list, the Fugaku uses only the power of its Fujitsu A64FX processors, and lacks accelerators…

Ultimately, the ISA is not always the most important when it comes to performance and consumption, but the microarchitecture , that is, the way in which that ISA is implemented. And, as this Green500 list has taught us, heterogeneous computing is a big plus .

What is Graph500?

Finally, there is another list or classification of HPC systems, but in this case it is not oriented to calculation performance like the Top500 or to energy efficiency like Green500, but rather it is oriented to position supercomputers according to intensive data loads. T this list is called Graph500 .

For those who don’t know, data-intensive computing refers to a type of parallel applications that are used to process large volumes of data, generally on the order of Terabytes or Petabytes in a matter of seconds, as occurs in Big Data. These types of applications are computationally intensive, and also I/O demanding.

The Graph500 was first published in 2010 at the ACM/IEEE Supercomputing Conference. This list is also based on a twice-yearly update, like the Top500, in June and in November, coinciding with each other.

Instead, instead of using the LINPACK benchmark in this case, this Graph500 uses performance metrics based on GTEPS (GigaTEPS) , where TEPS stands for Traversed Edges Per Second. This is a measure of communication capability, and contrasts with Floating Point Per Second (FLOPS) which is pure double-precision (64-bit) floating-point computing performance, or FP64, and does not take into account capabilities. machine communication.

Of course, GTEPS is not always needed in a processing center if the supercomputer is not going to be used for purposes such as Big Data , since, for example, CFD simulations or scientific applications often need FLOPS performance.

Thanks to Graph500, the most powerful supercomputers in the world can be cataloged according to these terms and thus be able to analyze the problems related to complex data . And it is that these systems need an emphasis on communication systems. To measure this performance in GTEPS, various programs are used as benchmarks , such as GNU Octave, versions of C programs using OpenMP, MPI, etc.

To give you an idea of ​​this list, I show you the following table updated in November 2022, with the top 3 on the Graph500 :

As you can see, in this case the columns are different from those of Top500 and Green500. In this case we have the usual ones for the position, country and site, the type of machine in question, number of nodes and cores, and finally we have two special columns, with the scale of the problem and the GTEPS score achieved. As we can see, the Fugaku is once again in prominent positions, in this case as a leader.

You should know that there are other benchmarks for HPC, such as the HPCG (High Performance Conjugate Gradient) , although this benchmark proposed by Michael Heroux from Sandia National Laboratories, along with Piotr Luszczek and Jack Dongarra from the University of Tennessee have not been as relevant. . In this case, to test the effect of memory subsystem limitations and internal interconnects. But this is another story…