The 62nd edition of the World's 500 Most Powerful Computers ranking has been published. The new Aurora cluster, deployed at the U.S. Department of Energy's Argonne National Laboratory, took second place in the 62nd edition. The cluster features nearly 4.8 million processor cores (Xeon CPU Max 9470 52C 2.4GHz CPU, Intel Data Center GPU Max accelerator) and delivers 585 petaflops of performance, 143 petaflops more than the cluster that previously held second place. Aurora uses SUSE as its operating system. Linux Enterprise Server 15 SP4.
The Frontier cluster, located at the U.S. Department of Energy's Oak Ridge National Laboratory, retained first place. The cluster boasts 8.7 million processor cores (AMD EPYC 64C 2GHz CPU, AMD Instinct MI250X accelerator) and delivers 1.194 exaflops of performance, twice as much as the second-place cluster (while consuming less power). HPE Cray OS (SUSE edition) is used as the operating system. Linux Enterprise Server 15).
Third place in the ranking was taken by the Eagle cluster, launched this year by Microsoft for the Azure cloud. The cluster contains 1.12 million processor cores (Xeon Platinum 8480C 48C 2GHz CPU) and demonstrates peak performance of 561 petaflops. The cluster software is based on Ubuntu 22.04.
The Fugaku cluster, located at the RIKEN Institute of Physical and Chemical Research (Japan), moved up to fourth place. The cluster is built using ARM processors (158976 nodes based on the Fujitsu A64FX SoC, equipped with a 48-core Armv8.2-A SVE 2.2GHz CPU). Fugaku delivers 442 petaflops of performance and runs Red Hat Enterprise. Linux.
Fifth place is occupied by the LUMI cluster, located at the European Supercomputing Center (EuroHPC) in Finland and providing performance of 379 petaflops. The cluster is built on the same HPE Cray EX235a platform as the leader of the rating, but includes 2.2 million processor cores (AMD EPYC 64C 2GHz, AMD Instinct MI250X accelerator, Slingshot-11 network). HPE Cray OS is used as the operating system.
Sixth, seventh and tenth places are occupied by the Leonardo, Summit and Sierra clusters, which were previously presented in the rating, but the 8th and 9th places were taken by newcomers - the MareNostrum 5 ACC clusters (681 thousand Xeon Platinum 8460Y+ 40C 2GHz cores, 138 petaflops, RHEL 9.1), deployed in the Barcelona supercomputer center, and the Eos NVIDIA DGX SuperPOD (486 thousand Xeon Platinum 8480C 56C 3.8GHz cores, 121 petaflops, Ubuntu 22.04.3), launched by NVIDIA.
Regarding domestic supercomputers, the Chervonenkis, Galushkin, and Lyapunov clusters created by Yandex dropped from 27th, 46th, and 52nd places to 36th, 58th, and 64th places. These clusters are designed to solve machine learning problems and provide performance of 21.5, 16, and 12.8 petaflops, respectively. The clusters are powered by Ubuntu 16.04 and are equipped with AMD EPYC 7xxx processors and NVIDIA A100 GPUs: the Chervonenkis cluster has 199 nodes (193 thousand AMD EPYC 7702 64C 2GH cores and 1592 NVIDIA A100 80G GPUs), Galushkin - 136 nodes (134 thousand AMD EPYC 7702 64C 2GH cores and 1088 NVIDIA A100 80G GPUs), Lyapunov - 137 nodes (130 thousand AMD EPYC 7662 64C 2GHz cores and 1096 NVIDIA A100 40G GPUs).
The Christofari Neo cluster deployed by Sberbank dropped from 55th to 67th place. Christofari Neo runs NVIDIA DGX OS 5 (edition Ubuntu) and demonstrates a performance of 11.95 petaflops. The cluster has over 98 computing cores based on an AMD EPYC 7742 64C 2.25GHz CPU and comes with an NVIDIA A100 80GB GPU. Sberbank's second cluster (Christofari) moved from 96th to 119th place in the rankings over the past six months.
Two more domestic clusters also remain in the ranking: Lomonosov 2 - shifted from 329 to 370 place (in 2015, the Lomonosov 2 cluster occupied 31 place, and its predecessor Lomonosov in 2011 - 13 place) and MTS GROM - shifted from 395 to 433 place . Thus, the number of domestic clusters in the ranking has not changed and, as six months ago, is 7 systems (for comparison, in 2020 there were 2 domestic systems in the ranking, in 2017 - 5, and in 2012 - 12).
The most interesting trends:
- Distribution by number of supercomputers in different countries:
- USA: 161 (150 - six months ago). The total productivity is estimated at 32.8% of the entire rating productivity (six months ago - 45.8%);
- China: 104 (134). In total, Chinese clusters generate 20.8% of all productivity (six months ago - 8.9%);
- Germany: 36 (36). Total productivity - 7.2% (7.2%);
- Japan: 32 (33). Total productivity - 6.4% (14.8%);
- France: 23 (24). Total productivity - 4.6% (4.8%);
- UK: 15 (14);
- Italy: 12 (7);
- South Korea 12 (8);
- Netherlands: 10 (8);
- Canada 10(10);
- Brazil 9 (9);
- Saudi Arabia 7 (6);
- Russia 7 (7);
- Sweden 6 (6);
- Australia 6 (5);
- Norway: 4 (4).
- Taiwan: 5.
- Poland: 4 (3);
- Ireland 4 (5);
- India: 4 (4);
- Switzerland 3 (4);
- Finland: 3 (3).
- Singapore: 3 (3);
- Spain: 3.
- In the ranking of operating systems used in supercomputers, since November 2017, only Linux;
- Distribution by distributions Linux (in brackets - 6 months ago):
- 44.6% (47%) use systems based on Linux, but do not detail the distribution;
- 11% (16%) use CentOS;
- 12.6% (10.8%) - RHEL;
- 9.6% (9.2%) — Cray Linux;
- 7.8% (6.4%) — Ubuntu;
- 4.4% (4.6%) - SUSE;
- 2% (1.6%) — Rocky Linux;
- 1% (1.2%) — Alma Linux;
- 0.2% (0.2%) — Amazon Linux;
- 0.2% (0.2%) — Scientific Linux.
- The minimum performance threshold for entering the Top500 over 6 months was 2.02 petaflops (six months ago - 1.87 petaflops). Five years ago, only 272 clusters showed performance of more than a petaflop, six years ago - 138, seven years ago - 94). For Top100, the entry threshold increased from 6.3 to 7.89 petaflops, and for Top10 - from 61.44 to 94.64 petaflops.
- The total performance of all systems in the rating over 6 months increased from 5.2 to 7 exaflops (four years ago it was 1.650 exaflops, and six years ago - 749 petaflops). The system that closes the current ranking was in 454th place in the last issue.
- The general distribution of the number of supercomputers in different parts of the world is as follows: 169 supercomputers are in Asia (192 - six months ago), 171 in North America (160) and 143 in Europe (133), 10 in South America (9), 6 in Oceania (5) and 1 in Africa (1).
- As a processor base, Intel CPUs are in the lead - 67.6% (six months ago it was 69.8%), AMD is in second place with 28% (24.2%), and IBM Power is in third place with 1.4% (was 1.4%).
- 21.4% (six months ago 21.6%) of all used processors have 24 cores, 21% (18.8%) - 64 cores, 10.6% (12%) - 20 cores, 7.4% (7.2%) - 32 cores, 6.2% (7.2% ) - 16 cores, 6% (7%) - 18 cores, 5.2% (5.8%) - 28 cores, 5% - 48 cores, 3% (4.2%) - 12 cores. The total number of processor cores in all clusters of the rating increased over six months from 95.5 million to 106.3 million.
- 185 out of 500 systems (six months ago - 185) additionally use accelerators or coprocessors, while 155 (168) systems use NVIDIA chips, 25 (11) use AMD, 2 (2) use Intel Xeon Phi, 1 (1 ) - MN-Core, 1 (1) - Matrix-2000.
- Among cluster manufacturers, Lenovo took first place - 33.8% (six months ago 33.6%), Hewlett-Packard Enterprise took second place - 20.6% (20%), EVIDEN took third place - 9.6% (0%), followed by Inspur - 6.8% (8.6%), Dell EMC 6.4% (4.8%), NVIDIA 3.6% (3.2%), NEC 2.4% (2%), Fujitsu 2.4% (2.4%), Sugon 1.8% (4.6%), MEGWARE 1.4% (1.6%), Microsoft Azure - 1.2 (1.25%), IBM 1.2% (1.2%), Penguin Computing - 1.2% (1%), Huawei 0.4% (0.4%).
- In the current ranking, the number of systems that use InfiniBand technology to connect nodes has for the first time outstripped the number of Ethernet-based systems. InfiniBand is used to connect nodes in 43.8% (six months ago 40%) of clusters, Ethernet is used in 41.8% (45.4%) of clusters, Omnipath - 6.6% (7%). Looking at overall performance, InfiniBand-based systems account for 41.4% (35.3%) of the Top500's overall performance, while Ethernet accounts for 44% (45.5%).
In the near future, a new release of the Graph 500 alternative rating of cluster systems is expected to be published, focused on evaluating the performance of supercomputer platforms associated with simulating physical processes and tasks for processing large amounts of data inherent in such systems. The Green500, HPCG (High-Performance Conjugate Gradient) and HPL-AI rankings are combined with the Top500 and reflected in the main Top500 ranking.
Source: opennet.ru
