55th edition 500 most powerful computers in the world. The June rating was headed by a new leader - the Japanese cluster , notable for the use of ARM processors.
Fugaku cluster at the Institute of Physical and Chemical Research and provides a performance of 415.5 petaflops, which is 2.8 more than the leader of the previous rating forced out to second place. The cluster includes 158976 nodes based on SoC , equipped with a 48-core CPU Armv8.2-A SVE (512 bit SIMD) with a clock frequency of 2.2GHz. In total, the cluster has more than 7 million processor cores (three times more than the leader of the previous rating), almost 5 PB of RAM and 150 PB of shared storage based on the Luster FS. The operating system is Red Hat Enterprise Linux.
Places from the second to the fifth were assigned to the previous leaders, who in the November ranking took from the first to the fourth places:
- Second place - cluster 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.
- Third place - American cluster , 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).
- Fourth place - Chinese cluster , 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.
- Fifth place - Chinese cluster , which includes almost 5 million cores and demonstrates a performance of 61 petaflops.
New clusters broke into sixth and seventh places (Italy, Dell EMC, 35 petaflops, 669k cores) and (USA, 27 petaflops, 277 thousand cores), which forced out the American cluster (Dell EMC, 23 petaflops, 448 thousand cores) to eighth place. The ninth place was taken by the new Italian cluster (IBM, 21.6 petaflops, 347 thousand cores), and the tenth is the Swiss cluster (Cray / HPE, 21.2 petaflops, 387 thousand cores), which in the previous ranking took 6th place.
The most interesting trends:
- Domestic cluster (6.6 petaflops, Ubuntu 18.04.01, built by Sberbank on the NVIDIA DGX-2 platform, uses a Xeon Platinum 8168 24C 2.7GHz CPU and has 99600 cores) moved from 6th to 29th place in the ranking in 36 months. Another domestic cluster Lomonosov 2 moved from 107th to 131st place. Cluster in , which occupied 465th place, was pushed out of the ranking. Thus, the number of domestic clusters in the ranking for six months decreased from 3 to 2 (in 2017, the ranking was 5 , and in 2012 - 12);
- Distribution by number of supercomputers in different countries:
- China: 226 (228 - six months ago). In total, Chinese clusters generate 45.2 of all productivity (six months ago - 31.9%);
- USA: 114 (117). The total productivity is estimated at 22.8% (six months ago - 37.8%);
- Japan: 29 (29);
- France: 19 (18);
- Germany: 16 (16);
- Netherlands: 15 (15);
- Ireland: 14 (14);
- Canada 12(9);
- UK: 10 (11);
- Italy: 7 (5);
- Brazil: 4 (3);
- Singapore 4 (4);
- South Korea, Saudi Arabia, Norway: 3;
- Russia, India, Australia, UAE, Switzerland, Sweden, Finland, Taiwan: 2.
- In the rating of operating systems used in supercomputers, only Linux has remained for three years;
- Distribution by Linux distributions (in brackets - 6 months ago):
- 54.4% (49.6%) do not detail the distribution,
- 24.6% (26.4%) use CentOS,
- 6.8% (6.8%) - Cray Linux,
- 6% (4.8%) - RHEL,
- 2.6% (3%) - SUSE,
- 2.2% (2%) - Ubuntu;
- 0.2% (0.4%) - Scientific Linux
- The minimum performance threshold for entering the Top500 has grown from 6 to 1142 teraflops in 1230 months (last year, only 272 clusters showed performance over petaflops, two years ago - 138, three years ago - 94). For the Top100, the entry threshold has increased from 2570 to 2801 teraflops;
- The total performance of all systems in the ranking increased from 1.65 to 2.23 exaflops over the year (749 years ago it was 449 petaflops). The system that closes the current rating was 348 in the last issue, and XNUMX in the year before;
- The general distribution of the number of supercomputers in different parts of the world is as follows:
272 supercomputers are located in Asia (274 - six months ago),
126 in North America (129) and 96 in Europe (94), 4 in South America and 2 in Oceania (3); - As a processor base, Intel CPUs are in the lead - 93.8% (six months ago it was 94%), IBM Power is in second place - 2.6% (was 2.8%), AMD is in third - 2.2% (0.6%), ARM (Marvell ThunderX2) is in fourth place. and Fujitsu A64FX) - 0.8%, on the fifth SPARC64 - 0.2% (0.6%). For the first time in the ranking, a cluster on ARM processors is presented, which immediately took first place.
- 37.4% (six months ago 35.6%) of all used processors have 20 cores, 12.2% (13.8%) - 16 cores, 10.6% - 24 cores,
10.4% (11%) - 18 cores, 9.8% (11.2%) - 12 cores, 7% (7.8%) - 14 cores; - 145 out of 500 systems (six months ago - 144) additionally use accelerators or coprocessors, while 135 systems use NVIDIA chips (six months ago there were 135), 6 - Intel Xeon Phi (was 5), 1 - PEZY (1), in 1 uses hybrid solutions (was 1), 1 uses Matrix-2000 (1), 1 AMD Vega GPU (1);
- Among manufacturers of clusters, Lenovo is in first place - 36% (six months ago 34.8%), in second place
Sugon 13.6% (14.2%), Inspur is in third place - 12.8% (13.2%), Hewlett-Packard is fourth - 7.6% (7%), followed by Cray 7.2%, Atos - 5.2% (4.6%), Fujitsu 2.6% (2.6%), IBM 2.4 (2.6%), Dell EMC 2% (2.2%), NVIDIA 1.4% (1.2%), Huawei 1.4% (2%),
Penguin Computing - 1.2% (2.2%). Three years ago, the distribution among manufacturers was as follows: Hewlett-Packard 28.6% (22.4%), in second place - Lenovo 17% (18.4%), in third place - Cray 11.4% (11.2%), in fourth place - Sugon 9.2% (9.4%), on the fifth β IBM 5.4% (6.6%). - To connect nodes in 52.6% of clusters, Ethernet is used (six months ago 52%), InfiniBand - 30.4% (28%), Omnipath - 9.8% (10%). If we consider the total performance, then Ethernet-based systems cover 52.6% (29%) of the total performance of Top500, InfiniBand - 30.4% (40%), Omnipath - 9.8%.
At the same time, a new release of the alternative rating of cluster systems is available , focused on evaluating the performance of supercomputer platforms associated with the simulation of physical processes and tasks for processing large amounts of data, typical for such systems. Rating separately more and merged with the Top500 as energy efficiency is now in the main Top500 rating (based on the ratio of LINPACK FLOPS to power consumption in watts).
Source: opennet.ru