Published the 58th edition of the ranking of the 500 most high-performance computers in the world. In the new issue, the top ten has not changed, but 4 new Russian clusters are included in the rating.
19th, 36th and 40th places in the ranking were occupied by the Russian clusters Chervonenkis, Galushkin and Lyapunov, created by Yandex to solve machine learning problems and providing performance of 21.5, 16 and 12.8 petaflops, respectively. The clusters run 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 cores 64C 2GH and 1088 NVIDIA A100 80G GPUs), Lyapunov - 137 nodes (130 thousand AMD EPYC 7662 64C 2GHz cores and 1096 NVIDIA A100 40G GPUs).
On the 43rd place was a new cluster of Sberbank - Christofari Neo, running NVIDIA DGX OS 5 (Ubuntu edition) and demonstrating a performance of 11.9 petaflops. The cluster has more than 98 thousand cores based on AMD EPYC 7742 64C 2.25GHz CPU and comes with NVIDIA A100 80GB GPU. The previously implemented Christofari cluster of Sberbank moved from 61st to 72nd place in the ranking in half a year.
Two more domestic clusters also remain in the ranking: Lomonosov 2 has shifted from 199th to 241st place (in 2015 the Lomonosov 2 cluster occupied 31st place, and its predecessor Lomonosov in 2011 - 13th place) and MTS GROM - has shifted from 240th to 294th place . Thus, the number of domestic clusters in the ranking increased from 3 to 7 in six months (for comparison, in 2020 there were 2 domestic systems in the ranking, in 2017 - 5, and in 2012 - 12).
As for the overall rating, the Japanese Fugaku cluster, built using ARM processors, remains in first place. The Fugaku cluster is hosted by the RIKEN Institute for Physical and Chemical Research and provides 442 petaflops of performance. The cluster includes 158976 nodes based on the Fujitsu A64FX SoC equipped with a 48-core CPU Armv8.2-A SVE (512 bit SIMD) clocked at 2.2GHz. In total, the cluster has more than 7.6 million processor cores (three times more than the previous leader), 5 PB of RAM and 150 PB of shared storage based on the Luster FS. The operating system is Red Hat Enterprise Linux. The total length of optical cables used to connect the nodes is about 850 kilometers.
In second place is the Summit cluster, which is deployed by IBM at the Oak Ridge National Laboratory (USA). The cluster runs Red Hat Enterprise Linux, includes 2.4 million processor cores (using 22-core IBM Power9 22C 3.07GHz CPUs and NVIDIA Tesla V100 accelerators), which provide 148 petaflops of performance, which is almost three times less than the rating leader.
The third place is occupied by the American cluster Sierra, installed in the Livermore National Laboratory by IBM on the basis of a platform similar to Summit and demonstrating performance at the level of 94 petaflops (about 1.5 million cores). The operating system is Red Hat Enterprise Linux.
In fourth place is the Chinese cluster Sunway TaihuLight, operating in the national supercomputing center of China, which includes more than 10 million computing cores and shows a performance of 93 petaflops. Despite similar performance, the Sierra cluster consumes half as much power as the Sunway TaihuLight. The operating system uses its own Linux distribution RaiseOS.
In fifth place is the Perlmutter cluster, manufactured by HPE and housed at the National Energy Research Center in the United States. The cluster includes 761 thousand cores based on the AMD EPYC 7763 64C 2.45GHz CPU and provides 71 petaflops of performance. Cray OS is used as the operating system.
The most interesting trends:
- Distribution by number of supercomputers in different countries:
- China: 173 (188 six months ago). In total, Chinese clusters generate 17.5% of all productivity (six months ago - 19.4%);
- USA: 149 (122). The total productivity is estimated at 32.5% (six months ago - 30.7%);
- Japan: 32 (34);
- Germany: 26 (23);
- France: 19 (16);
- Netherlands: 11 (16);
- UK: 11 (11);
- Canada 11(11);
- Russia 7 (3);
- South Korea 7 (5)
- Italy: 6 (6);
- Saudi Arabia 6 (6);
- Brazil 5 (6);
- Sweden 4 (3);
- Poland 4 (4);
- Australia, India, Switzerland, Finland: 3.
- In the rating of operating systems used in supercomputers, only Linux has been left for four and a half years;
- Distribution by Linux distributions (in brackets - two years ago):
- 51.6% (49.6%) do not detail the distribution,
- 18% (26.4%) use CentOS,
- 7.6% (4.8%) - RHEL,
- 7% (6.8%) - Cray Linux,
- 5.4% (2%) - Ubuntu;
- 4% (3%) - SUSE,
- 0.2% (0.4%) - Scientific Linux
- The minimum performance threshold for entering the Top500 for 6 months increased from 1511 to 1649 teraflops (three years ago, only 272 clusters showed performance over petaflops, four years ago - 138, five years ago - 94). For Top100, the entry threshold has increased from 4124 to 4788 teraflops;
- The total performance of all systems in the ranking increased from 2.8 to 3 exaflops over the year (two years ago it was 1.650 exaflops, and five years ago it was 566 petaflops). The system that closes the current rating was 433rd in the last issue, and 401 the year before;
- The general distribution of the number of supercomputers in different parts of the world is as follows: 226 supercomputers are located in Asia (245 six months ago), 160 in North America (133) and 105 in Europe (113), 5 in South America (6), 3 in Oceania (2) and 1 in Africa (1);
- As a processor basis, Intel CPUs are in the lead - 81.6% (two years ago it was 94%), AMD is in second place with 14.6% (0.6% !!), IBM Power is in third place - 1.4% (was 2.8%). There is an active increase in clusters based on AMD processors, for example, all new systems included in the Top15 are equipped with AMD CPUs.
- 26.6% (35.6 years ago 20%) of all used processors have 17.6 cores, 24% - 11.2 cores, 64% - 8.6 cores, 13.8% (16%) - 8.2 cores, 11% (18%) - 5.8 cores, 11.2% (12%) - XNUMX cores.
- 149 out of 500 systems (144 years ago - 143) additionally use accelerators or coprocessors, while 2 systems use NVIDIA chips, 5 - Intel Xeon Phi (was 1), 1 - PEZY (1), and XNUMX AMD Vega GPU ;
- Among cluster manufacturers, Lenovo is in first place - 36.8% (34.8% two years ago), Inspur is in second place - 11.6% (13.2%), Hewlett-Packard Enterprise is in third place - 9% (7%), followed by Sugon 7.8% (14.2%), Atos - 7.2% (4.6%), Cray 6.4% (7%), Dell EMC 3.2% (2.2%), Fujitsu 3% (2.6%), NVIDIA 2.4 (1.2%), NEC 2%, Huawei 1.4% (2%), IBM 1.4% (2.6%), Penguin Computing - 1.4% (2.2%). Seven years ago, the distribution among manufacturers looked like this: Hewlett-Packard 36%, IBM 35%, Cray 10.2% and SGI 3.8%;
- To connect nodes in 49.4% (two years ago 52%) of clusters, Ethernet is used, InfiniBand is used in 33.6% (28%) of clusters, Omnipath - 8.4% (10%). If we consider the total performance, then systems based on InfiniBand cover 43.3% of the total performance of the Top500, and Ethernet - 21.3%.
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