Jump to content

Fugaku (supercomputer)

From Wikipedia, the free encyclopedia
Fugaku
ActiveFrom 2021
SponsorsMEXT
OperatorsRiken
LocationRiken Center for Computational Science (R-CCS)
Architecture
Operating systemCustom Linux-based kernel
MemoryHBM2 32 GiB/node
Storage
  • 1.6 TB NVMe SSD/16 nodes (L1)
  • 150 PB shared Lustre FS (L2)[1]
  • Cloud storage services (L3)
Speed442 PFLOPS (per TOP500 Rmax), after upgrade; higher 2.0 EFLOPS on a different mixed-precision benchmark
CostUS$1 billion (total programme cost)[2][3]
RankingTOP500: No. 4, June 2024
PurposeScientific research
LegacyTOP500 No.1, June 2020 – June 2022
Websitewww.r-ccs.riken.jp/en/fugaku
SourcesFugaku System Configuration
PRIMEHPC FX1000 (Fugaku node) at SC19

Fugaku (Japanese: 富岳) is a petascale supercomputer at the Riken Center for Computational Science in Kobe, Japan. It started development in 2014 as the successor to the K computer[4] and made its debut in 2020. It is named after an alternative name for Mount Fuji.[5]

It became the fastest supercomputer in the world in the June 2020 TOP500 list[6] as well as becoming the first ARM architecture-based computer to achieve this.[7] At this time it also achieved 1.42 exaFLOPS using the mixed fp16/fp64 precision HPL-AI benchmark. It started regular operations in 2021.[8]

Fugaku was superseded as the fastest supercomputer in the world by Frontier in May 2022.[9]

Hardware

[edit]

The supercomputer is built with the Fujitsu A64FX microprocessor. This CPU is based on the ARM version 8.2A processor architecture, and adopts the Scalable Vector Extensions for supercomputers.[10] Fugaku was aimed to be about 100 times more powerful than the K computer (i.e. a performance target of 1 exaFLOPS).[11][12]

The initial (June 2020) configuration of Fugaku used 158,976 A64FX CPUs joined using Fujitsu's proprietary torus fusion interconnect.[7] An upgrade in November 2020 increased the number of processors.[13]

Software

[edit]

Fugaku uses a "light-weight multi-kernel operating system" named IHK/McKernel. The operating system uses both Linux and the McKernel light-weight kernel operating simultaneously and side by side. The infrastructure that both kernels run on is termed the Interface for Heterogeneous Kernels (IHK). The high-performance simulations are run on McKernel, with Linux available for all other POSIX-compatible services.[14][15][16]

Fugaku uses a three-tiered storage system to provide parallel storage to the compute nodes. The first-level LLIO storage is an NVM-based file I/O accelerator co-developed by Fujitsu and RIKEN that is allocated on per-job storage basis to the compute nodes for storing temporary data with low latency. The LLIO system stages data in and out of a second-level Fujitsu Exascale File System (FEFS), which uses disk-based storage based on Lustre software to provide a large persistent high-performance filesystem,[17][18] and a tape-based archive to store a large volume of infrequently accessed data.

Besides the system software, the supercomputer has run many kinds of applications, including several benchmarks. Running the mainstream HPL benchmark, used by TOP500, Fugaku is at petascale and almost halfway to exascale. Additionally, Fugaku has set world records on at least three other benchmarks, including HPL-AI; at 2.0 exaflops, the system has exceeded the exascale threshold for the benchmark.[19] A description of that benchmark is as follows:

The solver method of choice is a combination of LU factorization and iterative refinement performed afterwards to bring the solution back to 64-bit accuracy. The innovation of HPL-AI lies in dropping the requirement of 64-bit computation throughout the entire solution process and instead opting for low-precision (likely 16-bit) accuracy for LU, and a sophisticated iteration to recover the accuracy lost in factorization.[20]

Performance

[edit]

The reported initial performance of Fugaku was a Rmax of 416 petaFLOPS in the FP64 high performance LINPACK benchmark used by the TOP500.[7] After the November 2020 upgrade in the number of processors, Fugaku's performance increased to a Rmax of 442 petaFLOPS.[13]

In 2020, Fugaku also attained top spots in other rankings that test computers on different workloads, including Graph500, HPL-AI, and HPCG benchmark. No previous supercomputer has ever led all four rankings at once.[21]

After a hardware upgrade, as of November 2020, "Fugaku increased its performance on the new mixed precision HPC-AI benchmark to 2.0 exaflops, besting its 1.4 exaflops mark recorded six months ago. These represent the first benchmark measurements above one exaflop for any precision on any type of hardware." (a 42% increase)[22] Interestingly, the Arm A64FX core-count was only increased by 4.5%, to 7,630,848, but the measured performance rose much more on that benchmark (and the system does not use other compute capabilities, such as GPUs), and a little more on TOP500, or by 6.4%, to 442 petaflops, a new world record[23] and widening the gap to the next computer by that much. For the HPCG benchmark, it is 5.4 times faster, at 16.0 HPCG-petaflops, than the number two system, Summit,[24] which happens to also be second on TOP500.

Fugaku's performance surpasses the combined performance of the next 4 supercomputers on the TOP500 list (almost next 5) and surpasses by a 45% margin all the other top-10 computers on the HPCG benchmark.[25]

It has been reported since at least before May 2021 that China had developed a supercomputer with its own technology that exceeds the performance of Fugaku by a factor of two or more, and was secretly operating it to avoid sanctions by the United States.[26] In fact, a paper using this machine won the Gordon Bell Prize for best paper. As of May 2022, China is reportedly operating two supercomputers with exascale performance.[26]

History

[edit]

On May 23, 2019, Riken announced that the supercomputer was to be named Fugaku.[27] In August 2019, the logo for Fugaku was unveiled; it depicts Mount Fuji, symbolising "Fugaku's high performance" and "the wide range of its users".[4][28] In November 2019, the prototype of Fugaku won first place in the Green500 list.[29][30] Shipment of the equipment racks to the Riken facility began on December 2, 2019,[31] and was completed on May 13, 2020.[32] In June 2020, Fugaku became the fastest supercomputer in the world in the TOP500 list, displacing the IBM Summit.[7]

Fugaku has been used for research on masks related to the COVID-19 pandemic.[33][34]

In 2023, Fugaku has been used to develop Japanese language Generative AI Models by the Tokyo Institute of Technology, Riken Research Institute, Fujitsu Ltd and Tohoku University.[35]

Cost

[edit]

In 2018, Nikkei reported the programme would cost ¥130 billion (c. US$1 billion).[3][8]

Comparison

[edit]
Performance and cost comparison chart against computers ranked #1 in TOP500
Name Start year End year Performance
(PFLOPS)[note 1]
Cost
(million USD)

(not inflation adjusted)

TOP500 ranking CPU/GPU vendor CPU OS
Fugaku 2020 442[36] 1213[3][note 2] June 2020 to November 2021 1st[36] Fujitsu A64FX Linux (RedHat 8) and McKernel
Summit 2018 148 300[37] June 2018 to November 2019 1st IBM, Nvidia POWER9, Tesla Linux (RedHat)
Sierra 2018 94 November 2018 to November 2019 2nd
Sunway TaihuLight 2016 93 280[38] June 2016 to November 2017 1st NRCPC Sunway SW26010 Linux (Raise)
K 2011 2019 10 1045[39] June 2011 – November 2011 1st Fujitsu SPARC64 VIIIfx Linux

Those numbers are outdated. Today Fugaku ranks 4th in the TOP500.

  1. ^ This is based on the TOP500 Rmax measurement using the LINPACK benchmark at FP64 precision.
  2. ^ This is the total programme cost involving the creation of the technologies, rather than just for the machine.

See also

[edit]

References

[edit]
  1. ^ "Post-K (Fugaku) Information". Fujitsu. Archived from the original on 2020-06-08. Retrieved 2020-06-23.
  2. ^ Clark, Don (22 June 2020). "Japanese Supercomputer Is Crowned World's Speediest". The New York Times. Retrieved 26 June 2020.
  3. ^ a b c Takei, Tomohisa (14 September 2018). "お値段1300億円のポスト「京」、IT業界は今度こそ生かせるか". 日経クロステック(xTECH) (in Japanese). Nikkei. Retrieved 28 June 2020.
  4. ^ a b "スーパーコンピュータ「富岳」プロジェクト" (in Japanese). 理化学研究所. Retrieved 2019-05-25.
  5. ^ "Supercomputer Fugaku, named after Mt. Fuji, makes its debut". The Asahi Shimbun. 2020-06-16. Retrieved 2020-06-23.
  6. ^ "Japan's Fugaku gains title as world's fastest supercomputer" (Press release). www.riken.jp. Retrieved 2020-12-07.
  7. ^ a b c d Cutress, Dr Ian (22 June 2020). "New #1 Supercomputer: Fujitsu's Fugaku and A64FX take Arm to the Top with 415 PetaFLOPs". www.anandtech.com. Retrieved 22 June 2020.
  8. ^ a b Tsukimori, Osamu (7 January 2021). "Japan's Fugaku supercomputer is tackling some of the world's biggest problems". The Japan Times. Retrieved 26 January 2021.
  9. ^ "ORNL's Frontier First to Break the Exaflop Ceiling". Top500. 30 May 2022. Retrieved 2022-05-30.
  10. ^ "ポスト「京」のCPUの仕様を公表" (in Japanese). 富士通. 2018-08-22. Retrieved 2019-05-25.
  11. ^ "スパコン「京」後継機は「富岳」 計算性能100倍、21年稼働". 毎日新聞 (in Japanese). 2019-05-23. Retrieved 2019-05-30.
  12. ^ "Fugaku Remakes Exascale Computing In Its Own Image". 2019-12-09. Retrieved 2020-06-22.
  13. ^ a b "November 2020 summary – TOP500". www.top500.org. Retrieved 7 December 2020.
  14. ^ "Outline of the Development of the Supercomputer Fugaku". Riken Center for Computational Science. Archived from the original on 2021-01-23. Retrieved 2020-06-23.
  15. ^ "McKernel". Riken. Archived from the original on 2020-06-23. Retrieved 2020-06-23.
  16. ^ mckernel on GitHub
  17. ^ "Supercomputer Fugaku" (PDF). Fujitsu.
  18. ^ "Fugaku Codesign Report" (PDF). RIKEN Center for Computational Science.
  19. ^ "Fugaku Retains Title as World's Fastest Supercomputer". HPC Wire. 17 November 2020. Retrieved 27 November 2020.
  20. ^ "HPL-AI Mixed-Precision Benchmark — HPL-AI 0.0.2 documentation". icl.bitbucket.io. Retrieved 2020-11-18.
  21. ^ Byford, Sam (2020-06-23). "ARM-based Japanese supercomputer is now the fastest in the world". The Verge. Retrieved 2020-06-23.
  22. ^ "November 2020 | TOP500". www.top500.org. Retrieved 2020-11-18.
  23. ^ "Supercomputer Fugaku – Supercomputer Fugaku, A64FX 48C 2.2GHz, Tofu interconnect D | TOP500". www.top500.org. Retrieved 2020-11-18.
  24. ^ "HPCG – November 2020 | TOP500". www.top500.org. Archived from the original on 2021-05-06. Retrieved 2020-11-18.
  25. ^ "HPCG – November 2020 | TOP500". www.top500.org. Archived from the original on 2021-05-06. Retrieved 2020-12-01.
  26. ^ a b "US rushes to catch up with China in supercomputer race". Financial Times. 2022-05-18. Retrieved 2022-05-23.
  27. ^ "ポスト「京」の名称 「富岳(ふがく)」に決定" (in Japanese). 理化学研究所. 2019-05-23. Retrieved 2019-05-25.
  28. ^ "R-CCS announced the Fugaku logo | Riken Center for Computational Science Riken Website". www.r-ccs.riken.jp. Riken Center for Computational Science. Retrieved 23 June 2020.
  29. ^ "November 2019". TOP500.org. Retrieved 2019-11-20.
  30. ^ "Fugaku prototype named greenest supercomputer". Riken. 2019-11-18. Retrieved 2019-11-20.
  31. ^ "Fujitsu Begins Shipping Supercomputer Fugaku". Fujitsu. 2019-12-02. Retrieved 2020-06-23.
  32. ^ "Delivery of the Supercomputer Fugaku has been Completed". Riken Center for Computational Science. 2020-05-13. Retrieved 2020-06-23.
  33. ^ McCurry, Justin (26 August 2020). "Non-woven masks better to stop Covid-19, says Japanese supercomputer". The Guardian. Retrieved 22 September 2020.
  34. ^ McCurry, Justin (22 September 2020). "Face shields ineffective at trapping aerosols, says Japanese supercomputer". The Guardian. Retrieved 22 September 2020.
  35. ^ "Fugaku Supercomputer to Train AI in Japanese Language Models". The Japan News. 2023-05-23.
  36. ^ a b "TOP500 November 2021". Top500. June 2021. Retrieved 2021-11-16.
  37. ^ Shankland, Stephen (2018-06-26). "スパコン「TOP500」、IBM製「Summit」で米が中国を抜き首位に返り咲き" (in Japanese). ZDNet Japan. Retrieved 2020-10-28.
  38. ^ 伊本貴士 (2020-06-24). "頂上極めた「富岳」の次の挑戦、日本が強い分野の開発に生かせるか" (in Japanese). 日経クロステック. Retrieved 2020-10-28.
  39. ^ 田中誠士 (2019-08-05). "「2位じゃダメ」のスパコン京、見納め 6年超す長寿で" (in Japanese). 朝日新聞デジタル. Retrieved 2020-10-28.
[edit]


Records
Preceded by
IBM Summit
148.6 petaFLOPS
World's most powerful supercomputer
June 2020 – May 2022
0.54 exaFLOPS
Succeeded by
HPE Frontier
1.1 exaFLOPS