Fugaku – Fastest Supercomputer in the world today. Supercomputer Fugaku jointly developed by RIKEN and Fujitsu ranked No. 1 in the 55th TOP500 list of the world’s supercomputers with LINPACK performance of 415.53 PFLOPS (petaflops) and a computing efficiency ratio of 80.87%.
Fugaku also took first place in the international ranking HPCG (High-Performance Conjugate Gradient) achieving a high score of 13,400 TFLOPS (teraflops), while claiming 1st in the HPL-AI ranking with 1.421 EFLOPS.
The record achievement indicates the overall high performance of Fugaku and its significant contribution to make Society 5.0 a reality, leading Japan’s growth and producing world-leading results by solving social and scientific issues in the 2020s.
On June 22, 2020, Fujitsu announced that a supercomputer jointly developed by RIKEN and Fujitsu was ranked No. 1 in the 55th TOP500 list of the world’s supercomputers. Fugaku also took the No.1 position in the international ranking HPCG (High-Performance Conjugate Gradient), which measure the processing speed of the conjugate gradient method often used in practical applications including in the field of industry, and in the ranking of HPL-AI, which measures the performance of low-precision computing often used in AI such as deep learning.
These rankings were announced on June 22 at the ongoing virtual event ISC (International Supercomputing Conference) High Performance 2020 Digital.
Fugaku has been named after an alternative name to Mount Fuji. Fugaku has claimed that it is an exascale supercomputer developed by RIKEN and Fujitsu. RIKEN Center in Kobe, Japan is a large scientific research institute in Japan founded in 1917. The development of Fugaku was started in 2014 and it is the successor to the K computer. Its test operations have been started in 2021. Fugaku announced its launch in 2020 and became the fastest supercomputer in the world on the TOP500 list. The primarily ARM architecture-based computer, in June 2020, achieved 1.42 exaFLOPS (fp16 with fp64 precision) in the HPL-AI benchmark. Which makes it the first-ever supercomputer that achieved 1 exaOPS. As of March 2021, Fugaku is the fastest supercomputer in the world.
[email-subscribers-form id=”1″]
Some specs of Fugaku:
- The sponsor of Fugaku is MEXT
- Operator of Fugaku is RIKEN
- Architecture is based on Fujitsu A64FX CPU exists 48+4 cores per node which includes almost 158,976 nodes, Tofu interconnect D
- A memory of 32 GiB HBM2/node
- Storage L1: 1.6 TB NVMe SSD/16 nodes, L2: shared 150 PB Luster FS, L3: Cloud storage services
- Speed of 442 PFLOPS and after upgrade higher 2.0 EFLOPS benchmark
History behind Fugaku
On May 23, 2019, RIKEN decided the supercomputer is to be named Fugaku. In August 2019, the emblem for Fugaku got unveiled; it depicts Mount Fuji, symbolizing “Fugaku’s excessive-performance” and “the extensive variety of its users”. In November 2019, the prototype of Fugaku received the first position on the Green500 list. Shipment of the device racks to the RIKEN facility started on December 2, 2019, and finished on May 13, 2020. In June 2020, Fugaku has become the quickest supercomputer along with the global in the TOP500 list, displacing the IBM Summit. Also, Fugaku has been used for studies on masks associated with the COVID-19 pandemic.
Let’s see about its Software, Hardware, and Performance
Software
A lightweight multi-kernel operating machine named IHK/McKernel is used in the Fugaku. The OS makes use of both Linux and the McKernel lightweight kernel running concurrently and also has side-by-side usage. The infrastructure that each kernel runs on is called the Interface for Heterogeneous Kernels (IHK). The high-overall performance simulations operate on McKernel, with Linux for all other POSIX-well-suited services.
Besides the machine software, the supercomputer has run many types of applications, inclusive of numerous benchmarks. After running the mainstream HPL benchmark, The Fugaku is found to be at petascale and almost midway to exascale. Additionally, Fugaku has set international data as a minimum of 3 separate criteria, inclusive of HPL-AI; at 2.0 exaflops, the machine has surpassed the exascale threshold for the benchmark.
Also Read: China’s Jiuzhang achieves Quantum Supremacy with light.
Hardware
The supercomputer is developed with the Fujitsu A64FX microprocessor. This CPU is primarily based totally on the ARM model 8.2A processor architecture and adopts the Scalable Vector Extensions for supercomputers. Fugaku’s efficiency seemed to be approximately one hundred instances greater effective than the K computer (i.e. an overall performance goal of 1 exaFLOPS).
The initial (June 2020) configurations of Fugaku used 158,976 A64FX CPUs joined collectively that make use of Fujitsu’s proprietary torus fusion interconnect. An improvement in November 2020 has been done that multiplied the number of processors.
Performance
Measurement Results of Fugaku
1. TOP500
The Fugaku system ranked first in the TOP500 list consisted of 396 racks (152,064 nodes, approximately 95.6% of the entire system), and the LINPACK performance was 415.53 PFLOPS (petaflops) with the computing efficiency ratio of 80.87%. It is the first time for a Japanese supercomputer to take first place in TOP500 since the K computer claimed No.1 in November 2011 (the 38th TOP 500 list). Fugaku’s performance is approximately 2.8 times that of the supercomputer ranked second in the TOP500 list with148.6 PFLOPS.
2. HPCG
For this benchmark, 360 racks (138,240 nodes, approximately 87% of the entire system) of Fugaku were used to achieve the high score of 13,400 TFLOPS (teraflops). This proves that the supercomputer can efficiently handle such real-world applications in the field of industry and perform well. Moreover, Fugaku exceeds the performance of the No.2 supercomputer (2,925.75 TFLOPS) by approximately 4.6 times.
3. HPL-AI
Unlike the conventional listings of TOP500 and HPCG which measure the performance of the double-precision arithmetic logic unit, HPL-AI is a new benchmark established in November 2019 as an index for evaluating calculation performance that takes into account the capabilities of single-precision and half-precision arithmetic logic units used in artificial intelligence. For this measurement, a high score of 1.421 EFLOPS (Exa FLOPS) was recorded using 330 racks (126,720 nodes, approximately 79.7% of the entire system) of Fugaku.
This is also a historical record, as Fugaku achieved 1 exa (10 raised to the power of 18) in one of the HPL benchmarks for the first time in the world. This proves Fugaku’s capability to contribute to the advancement of Society 5.0, as a research platform for machine learning and big data analysis.
Some comparisons among the top 3 Supercomputers of the top100 list:
1. Fugaku
- It was introduced in the year 2020
- CPU is based on A64FX
- GPU vendor is Fujitsu
- OS is based on Custom Linux-based Kernel
- Performance according to PFLOPS is 415
- It got ranked at the top500 list on 1st June 2020
- And costs around $1213 million
2. Summit
- It was introduced in the year 2018
- CPU is based on Tesla and POWER9
- GPU vendor is IBM and NVIDIA
- OS is based on Linux RedHat
- Performance according to PFLOPS is 148
- It got ranked at the top500 list between 1st, June 2018 and November 2019
- And costs around $300 million
3. Sierra
- It was introduced in the year 2018
- CPU is based on Tesla and POWER9
- GPU vendor is IBM and NVIDIA
- OS is based on Linux RedHat
- Performance according to PFLOPS is 94
- It got ranked at the top500 list between 2nd, November 2018 and 2019
- And costs around $300 million
Discussions
Satoshi Matsuoka, Director, Riken-Center for Computational Science (R-CCS) who developed Fugaku – Fastest Supercomputer in the world said “Ten years after the initial concept was proposed, and six years after the official start of the project, Fugaku is now near completion. Fugaku was developed based on the idea of achieving high performance on a variety of applications of great public interest, such as the achievement of Society 5.0, and we are very happy that it has shown itself to be outstanding on all the major supercomputer benchmarks. In addition to its use as a supercomputer, I hope that the leading-edge IT technology developed for it will contribute to major advances on difficult social challenges such as COVID-19”
Naoki Shinjo, Corporate Executive Officer, Fujitsu Limited said “I believe that our decision to use a co-design process for Fugaku, which involved working with RIKEN and other parties to create the system, was a key to our winning the top position on a number of rankings. I am particularly proud that we were able to do this just one month after the delivery of the system was finished, even during the COVID-19 crisis. I would like to express our sincere gratitude to RIKEN and all the other parties for their generous cooperation and support. I very much hope that Fugaku will show itself to be highly effective in real-world applications and will help to make Society 5.0 a reality.”
About the supercomputer benchmarks
1. TOP500
The TOP500 list is a project that regularly ranks and evaluates the top 500 fastest supercomputer systems in the world based on LINPACK performance. Developed by Dr. Jack Dongarra of the University of Tennessee, US, to solve a system of linear equations by matrix calculation, the LINPACK program was launched in 1993 to announce the supercomputer ranking two times a year (June and November).
LINPACK measures the computing power of double-precision floating-point numbers used in many scientific and industrial applications and to get a high score on this benchmark, it is necessary to run a large-scale benchmark for a long time. In general, a high LINPACK score is said to be a comprehensive measure of computing power and reliability.
2. HPCG
The TOP500 has long been a popular benchmark for evaluating computing power, which was an important performance indicator for solving a system of linear equations composed of a dense coefficient matrix. More than 20 years have passed since the project was launched in 1993, and recently it has been pointed out that the performance requirements of actual applications are not met, and the time required for benchmark testing is prolonged.
Accordingly, Dr. Dongarra et al. proposed a new benchmark program, HPCG, that uses the conjugate gradient method to solve a system of linear equations composed of a sparse coefficient matrix, which are often used in industrial applications. Following the announcement of measurements on the world’s leading 15 supercomputer systems at ISC 2014 in June, the official ranking was announced at the International Conference for High-Performance Computing, Networking, Storage, and Analysis (SC14) held in New Orleans, the US, in November.
3. HPL-AI
The TOP500 and HPCG have ranked supercomputers in terms of computational performance for solving a system of linear equations. In both cases, it was stipulated in the rules that only double-precision arithmetic (16-digit floating-point number in 10), which has been widely used in scientific and technological calculations as well as industrial applications, should be used for calculations.
In recent years, more computers, equipped with GPUs or AI dedicated chips, are adding a large number of low-precision arithmetic logic units (5 or 10 digits in 10) to increase their performance. Since these high-performance computing capabilities are not reflected in the TOP500 list, Dr. Dongarra et al. improved the LINPACK benchmark by allowing the use of low precision calculations and proposed a new benchmark, HPL-AI, in November 2019.
HPL-AI allows LINPACK to perform low-precision computations when solving a system of linear equations using LU decomposition. However, since the calculation accuracy is inferior to that of double-precision calculation, it is required to obtain the same accuracy as double-precision calculation by a technique called iterative refinement. In other words, it’s a two-step benchmark. As the HPL-AI rules were issued in November 2019, this is the first announcement of the benchmark ranking.
RIKEN couldn’t use the system’s complete energy for the June competition due to the fact his crew didn’t have sufficient time, Matsuoka explained that they only had about two weeks in their hand to the closing date from the time they added up the very last nodes of the machine. They had little or no time, and some benchmarks were compromised. Since then, his crew had enough possibility now no longer most effective to convey up the closing nodes withinside the complete cluster however additionally to tune the code for optimum performance.
Fugaku – Fastest Supercomputer in the world scored what the supercomputing enterprise now calls a “triple,” scoring first at the accompanying HPCG and HPL-AI benchmark tests. HPCG affords what long-time Top500 co-maintainer Martin Meuer called “every other attitude into the hardware,” giving greater favorable rankings to structures that are tuned for performance and manipulate of reminiscence bandwidth, for instance. HPL-AI is a miles quicker jogging test, buying and selling high-precision floating-factor operations for lower-precision math greater generally utilized in device learning.
Fugaku’s prolonged overall performance lead represents even extra achievement for Arm architectures, which have been receiving lots extra scrutiny now that the bodily obstacles of Moore’s Law seem to were reached. Data Center Knowledge (DCK) requested Matsuoka if there are any instructions to be discovered from Fujitsu’s improvement of its supercomputer-unique processors which could explain why Fugaku keeps reveling in the sort of overall performance benefit over x86-primarily based structures at the list.
In the first lesson, he instructed DCK, which turned into the significance of co-design. “When, globally, they launched into this undertaking of looking to attain exascale (1 exaflop per second and beyond), The emphasis now lies on simply converting to the exascale, however clearly to excel in software overall performance.
The eventual co-designers of A64FX decided that, at its modern boom trajectory, reminiscence bandwidth have to be x86 server-magnificence CPUs might contact 200GB/s (gigabyte in keeping with second). For the programs RIKEN started planning, they might want 1 TB/s. Soon, AMD, Samsung, and SK Hynix began out taking part in High-Bandwidth Memory (HBM), wherein bandwidth can be accelerated via way of means of stacking DRAM modules atop one another. That maybe rating a few factors with cloud or organization server manufacturers, however, He said the appreciation of caching HBM hierarchies and chip interconnects could create problems; elements that might have been adverse to the deterministic overall performance that HPC additives require.
Research establishments are a good deal extra inclined to undertake new technologies, to strive out new things, and to go along with architectures that are not the least unusual place with the market. For example, IBM Power AC922 structures coupled with Nvidia GPUs (8 of the present-day Top500 structures), AMD-primarily based structures (24 of 500), and Arm-primarily based structures (3 of 500).