Supercomputer "Chervonenkis" in the Yandex data center in Sasov, Ryazan region.
Research Computing Center of Moscow State University named after M.V. Lomonosov and the Interdepartmental Supercomputing Center of the Russian Academy of Sciences published the 37th edition (dated September 26, 2022) of the list of the top 50 most powerful computers in the CIS. In the top five of the rating, there are three Yandex supercomputers (Chervonenkis, Galushkin and Lyapunov) and two Sberbanks (Christofari Neo and Christofari).
The total performance of all systems on the Linpack test increased from 85.1 petaflops to 85.4 petaflops in six months. The total peak performance of the list systems was 120.6 petaflops (120.2 petaflops in the previous edition of the list). In just six months, three new supercomputers appeared on the list, taking 20th place (FGBOU VO Mari State University), 39th place (Tula Industrial Technopark) and 45th place (Pangea PD Cluster).
The current edition of the top 50 list required 83.9 teraflops of performance on the Linpack test (77.5 teraflops in the previous edition).
In 41 out of 50 supercomputers of this edition, solutions from Intel are installed as the main processors. The number of hybrid supercomputers using accelerators for calculations has grown from 32 to 34 in six months.
Yandex supercomputers are named after Soviet and Russian scientists who contributed to machine learning theory and computer science. The Lyapunov system was put into operation in December 2020, and Chervonenkis and Galushkin in June 2021. Yandex supercomputers are based on AMD EPYC processors and Nvidia A100 graphics accelerators with InfiniBand interconnect based on Mellanox switches. They are based on the Nvidia HGX A100 architecture, optimized for the machine learning tasks facing Yandex. Using this architecture allowed system developers to increase the cluster size and train the largest ML models about twice as fast as with the standard architecture.
The most powerful of them is Chervonenkis, it has a peak performance of 21.53 petaflops. Galushkin has 16.02 petaflops, while Lyapunov has 12.81 petaflops. Yandex uses its supercomputers to train neural networks. For example, thanks to them, the Yandex Translator service translates texts, pictures and videos more accurately and faster, and Yandex Direct selects more relevant ads. For the Yandex Search service, language models from the YaLM (Yet another Language Model) family, inspired by GPT-3 from Open AI, help compose and rank quick responses. Yandex supercomputers taught the Alisa voice assistant to maintain a lively dialogue with users.
Currently, Chervonenkis, Galushkin and Lyapunov are ranked 22nd, 40th and 43rd, respectively, in the ranking of the top 500 supercomputers in the world. The Sberbank supercomputer Christofari Neo is in 46th place there, and Christofari is in 80th place. In total, only 7 systems from Russia are in the ranking of the top 500 supercomputers in the world.
On November 11, 2021, Sber introduced its second supercomputer, Christofari Neo. Its effective double precision performance is almost 12 petaflops (11.95 petaflops). The second model of the Sbera supercomputer was created on the basis of Nvidia technologies, it is based on Nvidia A100 graphics processors with 80 GB of memory, and more than 700 of them are installed there. Sber specified that the A100 tensor cores provide the highest performance in AI tasks, and 80 GB of memory allow you to work with large AI models and data arrays.
Sber's first supercomputer, Christofari, was launched at the end of 2019, also in partnership with Nvidia, based on high-performance Nvidia DGX-2 nodes equipped with Tesla V100 computing accelerators. At the beginning of its operation, the performance of the first Sber supercomputer in the tests carried out reached 6.669 petaflops (peak performance is 8.789 petaflops). Description of the composition of Christofari: Nvidia DGX-2, Xeon Platinum 8168 24C 2.7GHz, Mellanox InfiniBand EDR, Nvidia Tesla V100, working cores: 99,600.
The US sanctions imposed at the end of August to ban Nvidia and AMD from supplying high-performance chips to the Russian Federation and providing support for current supplies will affect the development and scaling of supercomputer systems and computing clusters of Yandex and Sber cloud services. The high-performance systems of these Russian companies are based on Nvidia technologies and equipment, including graphics accelerators with A100 and H100 tensor cores, which are now banned from delivery.Industry experts believe that companies that used the principle “if we don’t have enough computing power, buy some more Nvidia GPUs” will suffer from the ban on the export of Nvidia and AMD chips to Russia. A Yandex representative told the media that the company's services and IT infrastructure are now operating as normal. Representatives of Sberbank declined to comment.
In early July, Nvidia stopped selling and renewing licenses for special software for cloud gaming, which also affected solutions based on Nvidia GRID technology that are popular with corporate customers and Russian gaming services.
bbabo.Net