An buga bugu 60 na kimar mafi girman manyan kwamfutoci

An buga bugu na 60 na kima na kwamfutoci 500 mafi kyawun aiki a duniya. A cikin sabon bugu, akwai canji ɗaya kawai a cikin manyan goma - ƙungiyar Leonardo, wanda ke cikin cibiyar binciken kimiyyar Italiyanci CINECA, ya ɗauki matsayi na 4. Tarin ya haɗa da kusan nau'ikan sarrafawa miliyan 1.5 (CPU Xeon Platinum 8358 32C 2.6GHz) kuma yana ba da aikin petaflops 255.75 tare da amfani da wutar lantarki na kilowatts 5610.

Manyan ukun, daidai da watanni 6 da suka gabata, sun haɗa da gungu masu zuwa:

• Frontier - Yana zaune a Ma'aikatar Makamashi ta Amurka ta Oak Ridge Laboratory National. Tarin yana da kusan nau'ikan kayan sarrafawa miliyan 9 (AMD EPYC 64C 2GHz CPU, AMD Instinct MI250X accelerator) kuma yana ba da aikin exaflops 1.102, wanda kusan sau uku ya fi gungu na biyu (yayin da ikon Frontier ya kasance ƙasa da 30%).
• Fugaku - yana zaune a RIKEN Institute of Physical and Chemical Research (Japan). An gina gungu ta amfani da na'urori masu sarrafa ARM (158976 nodes dangane da Fujitsu A64FX SoC, sanye take da 48-core Armv8.2-A SVE 2.2GHz CPU). Fugaku yana ba da 442 petaflops na aiki.
• An shirya LUMI a Cibiyar Supercomputing ta Turai (EuroHPC) a Finland kuma tana ba da 151 petaflops na aiki. An gina gungu akan dandamalin HPE Cray EX235a iri ɗaya a matsayin jagoran ƙimar, amma ya haɗa da na'urorin sarrafawa miliyan 1.1 (AMD EPYC 64C 2GHz, AMD Instinct MI250X accelerator, Slingshot-11 network).

Dangane da manyan kwamfutoci na cikin gida, gungu na Chervonenkis, Galushkin da Lyapunov da Yandex ya kirkira sun ragu daga wurare 22, 40 da 43 zuwa wurare 25, 44 da 47. An tsara waɗannan gungu don magance matsalolin koyon inji da samar da aikin 21.5, 16 da 12.8 petaflops, bi da bi. Rukunin suna gudanar da Ubuntu 16.04 kuma an sanye su da na'urori masu sarrafawa na AMD EPYC 7xxx da NVIDIA A100 GPUs: gungu na Chervonenkis yana da nodes 199 (193 dubu AMD EPYC 7702 64C 2GH cores da 1592 NVIDIA A100 80G GPUs 136 NVDIA EPYC - 134G GPUs), Galush cores 7702 64C 2GH da 1088 GPU NVIDIA A100 80G), Lyapunov - 137 nodes (130 dubu cores AMD EPYC 7662 64C 2GHz da 1096 GPU NVIDIA A100 40G).

Tarin Christofari Neo da Sberbank ya tura ya ragu daga matsayi na 46 zuwa na 50. Christofari Neo yana gudanar da NVIDIA DGX OS 5 (bugu na Ubuntu) kuma yana nuna aikin 11.9 petaflops. Tarin yana da fiye da nau'ikan kwamfuta fiye da 98 dangane da AMD EPYC 7742 64C 2.25GHz CPU kuma ya zo tare da NVIDIA A100 80GB GPU. Tari na biyu na Sberbank (Christofari) ya tashi daga matsayi na 80 zuwa 87 a cikin matsayi sama da watanni shida.

Har ila yau, wasu gungu na gida biyu sun kasance a cikin matsayi: Lomonosov 2 - ya koma daga 262 zuwa 290 (a cikin 2015, Lomonosov 2 cluster ya dauki matsayi 31, kuma Lomonosov wanda ya gabace shi a 2011 - 13 wuri) da MTS GROM - daga 318 zuwa 352 wuri . Don haka, adadin gungu na gida a cikin martaba bai canza ba kuma, kamar yadda watanni shida da suka gabata, shine tsarin 7 (don kwatanta, a cikin 2020 akwai tsarin gida 2 a cikin martaba, a cikin 2017 - 5, kuma a cikin 2012 - 12).

Abubuwan da suka fi ban sha'awa:

• Rarraba ta adadin manyan kwamfutoci a ƙasashe daban-daban:
  • China: 162 (173 - watanni shida da suka wuce). A cikin duka, gungu na kasar Sin suna samar da kashi 10% na duk abin da ake samarwa (watanni shida da suka gabata - 12%);
  • Amurka: 127 (127). An kiyasta jimlar yawan aiki a 43.6% na duk yawan yawan ƙimar (watanni shida da suka gabata - 47.3%);
  • Jamus: 34 (31). Jimlar yawan aiki - 4.5%;
  • Japan: 31 (34). Jimlar yawan aiki - 12.8%;
  • Faransa: 24 (22). Jimlar yawan aiki - 3.6%;
  • Birtaniya: 15 (12);
  • Kanada 10 (14);
  • Netherlands: 8 (6);
  • Koriya ta Kudu 8 (6)
  • Brazil 8 (6);
  • Rasha 7 (7);
  • Italiya: 7 (6);
  • Saudi Arabia 6 (6);
  • Sweden 6 (5);
  • Ostiraliya 5 (5);
  • Ireland 5;
  • Poland 5 (5);
  • Switzerland 4 (4);
  • Finland: 3 (4).
  • Singapore: 3;
  • Indiya: 3;
  • Poland: 3;
  • Norway: 3.
• A cikin martabar tsarin aiki da ake amfani da su a cikin manyan kwamfutoci, Linux kawai ya rage tsawon shekaru shida;
• Rarraba ta rarrabawar Linux (a cikin baka - watanni 6 da suka gabata):
  • 47.8% (47.8%) ba su dalla-dalla rarraba;
  • 17.2% (18.2%) suna amfani da CentOS;
  • 9.6% (8.8%) - RHEL;
  • 9% (8%) - Cray Linux;
  • 5.4% (5.2%) - Ubuntu;
  • 3.8% (3.8%) - SUSE;
  • 0.8% (0.8%) - Alma Linux;
  • 0.8% (0.8%) - Rocky Linux;
  • 0.2% (0.2%) - Linux na Kimiyya.
• Matsakaicin iyakar aiki don shigar da Top500 na watanni 6 shine 1.73 petaflops (watanni shida da suka gabata - 1.65 petaflops). Shekaru hudu da suka gabata, gungu 272 ne kawai suka nuna aikin fiye da petaflop, shekaru biyar da suka gabata - 138, shekaru shida da suka gabata - 94). Don Top100, ƙofar shiga ya karu daga 5.39 zuwa 9.22 petaflops;
• Jimlar ayyukan duk tsarin a cikin ƙimar sama da watanni 6 ya ƙaru daga 4.4 zuwa 4.8 exaflops (shekaru uku da suka wuce 1.650 exaflops ne, kuma shekaru biyar da suka gabata - 749 petaflops). Tsarin da ya rufe matsayi na yanzu ya kasance a matsayi na 458 a cikin fitowar ta ƙarshe;
• Babban rabon manyan kwamfutoci a sassa daban-daban na duniya shine kamar haka: 218 supercomputers suna cikin Asiya (229 - watanni shida da suka gabata), 137 a Arewacin Amurka (141) da 131 a Turai (118), 8 a Kudancin Amurka. Amurka (6), 5 a Oceania (5) da 1 a Afirka (1);
• A matsayin tushen tsarin sarrafawa, Intel CPUs suna kan gaba - 75.6% (watanni shida da suka gabata shine 77.4%), AMD yana matsayi na biyu da 20.2% (18.8%), IBM Power yana matsayi na uku - 1.4% (ya kasance 1.4). %).
• 22.2% (watanni shida da suka wuce 20%) na duk na'urorin da aka yi amfani da su suna da nau'ikan 24, 15.8% (15%) - 64 cores, 14.2% (19.2%) - 20 cores, 8.4% (8.8%) - 16 cores, 7.6% 8.2%) - 18 cores, 6% - 28 cores, 5% (5.4%) - 12 cores.
• 177 cikin 500 tsarin (watanni shida da suka gabata - 167) kuma suna amfani da accelerators ko coprocessors, yayin da tsarin 161 ke amfani da kwakwalwan kwamfuta na NVIDIA, 9 - AMD, 2 - Intel Xeon Phi (daga 5), ​​1 - PEZY (1), 1 - MN- Core, 1 - Matrix-2000;
• Daga cikin masana'antun gungu, Lenovo ya ɗauki matsayi na farko tare da 32% (watanni shida da suka gabata 32%), Kamfanin Hewlett-Packard ya ɗauki matsayi na biyu tare da 20.2% (19.2%) Inspur ya ɗauki matsayi na uku tare da 10% (10%), Atos ya biyo baya - 8.6% (8.4%), Sugon 6.8% (7.2%), Dell EMC 3.6% (3.4%), NVIDIA 2.8% (2.8%), NEC 2.4% (2%), Fujitsu 2% (2.6%), MEGWARE 1.2 %, Penguin Computing - 1.2% (1.2%), IBM 1.2% (1.2%), Huawei 0.4% (1.4%).
• Ana amfani da Ethernet don haɗa nodes a cikin 46.6% (watanni shida da suka gabata 45.4%) na gungu, ana amfani da InfiniBand a cikin 38.8% (39.2%) na gungu, Omnipath - 7.2% (7.8%). Duban aikin gabaɗaya, tsarin tushen InfiniBand yana da kashi 33.6% (32.4%) na babban aikin Top500, yayin da Ethernet ke lissafin 46.2% (45.1%).

Nan gaba kadan, ana sa ran buga wani sabon bugu na madadin kima na tsarin tari mai lamba 500, wanda aka mayar da hankali kan tantance ayyukan dandali na manyan kwamfutoci da ke da alaƙa da simintin hanyoyin jiki da ayyuka don sarrafa bayanai masu yawa na irin waɗannan tsarin. Green500, HPCG (High-Performance Conjugate Gradient) da HPL-AI ana haɗe su tare da Top500 kuma suna nunawa a cikin babban ƙimar Top500.

source: budenet.ru