The performance of the “Govorun” supercomputer at JINR has reached 2.2 PFlops, opening up new possibilities for the application of machine learning methods in particle physics Пресс-релиз Dubna, 7 July 2025 – Specialists from the RSC Group of Companies in cooperation with the Meshcheryakov Laboratory of Information Technologies of the Joint Institute for Nuclear Research (MLIT JINR) have completed the next stage of modernization of the “Govorun” supercomputer. The computing resources of this system were expanded by installing two nodes on top of the new RSC Exastrim AI server solution with eight NVidia H100 graphics accelerators in each and direct liquid cooling. The new servers are unique and were designed and manufactured for the “Govorun” supercomputer taking into account its architectural features. At the same time, the peak performance of the GPU component of the “Govorun” supercomputer has grown by 36% and reached 1.4 PFlops (in double-precision operations), and the complex’s total capacity is now 2.2 PFlops in double-precision operations. “The RSC Exastrim AI servers, installed in the “Govorun” supercomputer, are a modern modular hardware and software platform for high-performance computing, aimed at solving scientific and engineering tasks using artificial intelligence technologies. We believe that this will facilitate the wider and more intensive use of various deep learning algorithms in JINR research programs,” MLIT Director Sergei Shmatov emphasized. The RSC Exastrim AI solution is a computing server with a pool of graphics cards (GPGPU) with direct liquid cooling, paired together by bridges on top of NVLink technology for high-speed connections to ensure fast data transfer between GPUs. In addition, it has a local subsystem for warm data storage and a network subsystem with access using GPUDirect technology. RSC Exastrim AI has the ability to enlarge resources by connecting additional pairs of GPUs or an external data storage system based on a pool of solid-state disks (JBOF), connected directly to the server. Configuration of the RSC Exastrim AI servers installed at JINR: node height: 2U, 2 Intel Xeon Platinum 8468 processors (4th generation, 48 cores, frequency 2.1–3.8 GHz, cache 105 MB), 8 NVIDIA H100 graphics accelerators (PCIe, 80 GB), 1 TB of RAM, NVMe SSD drives with 16 TB of storage capacity, 4 power supply units manufactured by the RSC, direct liquid cooling system developed by the RSC. Moreover, at the end of 2024, the data storage system (DSS) of the “Govorun” supercomputer was enhanced, resulting in an increase in the total storage capacity by 25%, from 8 to 10 PB. Two new RSC Tornado AFS data storage nodes with a capacity of 1 PB each were added to the JINR computing complex’s DSS. The RSC Tornado AFS DSS updated modification embraces a server board based on 4th generation Intel Xeon processors and a switch with a PCIe Gen4 interface, which made it possible to install two interconnect adapters with a throughput of 200 Gbps each. The RSC Tornado AFS DSS supports GPUDirect Storage (GDS) technology, which enables direct data transfer between local or remote storage and GPU memory. “The flexible configuration system implemented in the RSC Tornado AFS data storage system utilizing the RSC BasIS Storage module allows us to maximize the efficiency of using the “Govorun” supercomputer’s resources to solve a wide range of tasks the Institute is facing. These tasks have various requirements for the performance parameters of the data access subsystem, they may require both high speed for the output or input of large chunks of data and handling many requests to small chunks of data. In particular, two new RSC Tornado AFS servers are already used for the Lustre distributed file system to work with data from the NICA megascience project,” MLIT Scientific Leader Vladimir Korenkov pointed out. The “Govorun” computing system employs the RSC BasIS 4 integrated software package and the RSC BasIS Storage module (included in the Register of Russian Software). RSC BasIS Storage is a software module developed to automate the process of creating static and dynamic data storage systems at the request of users or operators of the computing complex. The functionality of this module is implemented using the RSC BasIS 4 software solution, which allows the operator of the computing system to describe all components of the cluster and establish static or dynamic links between control objects, i.e., configure the entire complex. The RSC BasIS 4 micro-agent architecture enables not only to maintain all system objects in working order, but also to actively interact with them, thereby enhancing the efficiency of utilizing computing resources. They are supplied to customers as part of software and hardware solutions on top of the RSC Tornado high-density and energy-efficient platform and as part of other RSC solutions. RSC BasIS, in combination with RSC hardware platforms, allows creating hyperconverged solutions for high-performance computing and the efficient processing of large data volumes. NICA megaproject to build an accelerator complex at JINR to recreate in laboratory conditions a special state of matter called quark-gluon plasma in which our Universe was in the first moments after the Big Bang. The implementation of this megaproject started in 2013. It is noteworthy that two data storage systems previously installed by RSC specialists as part of the “Govorun” supercomputer of the Joint Institute for Nuclear Research are included in the world IO500 ranking of the most high-performance data storage systems. The current IO500 list (June 2025) comprises a DSS segment on top of the RSC Storage on-Demand solution with support for a new open-source storage system, DAOS, it is the leader among Russian data storage systems and occupies the 139th position. This project, based on the DAOS distributed object storage system, is implemented for data processing within the NICA megaproject. A storage segment on top of the Lustre file system at JINR is currently ranked 167th in this list. It is worth reminding that the “Govorun” supercomputer DSS took a high 9th place in the world IO500 list in June 2018, immediately after installation and commissioning. The data storage system at JINR was able to occupy such a high position in this world ranking due to the application of a unique technology for creating multi-layer, scalable, distributed and composable on-demand data storage systems of the RSC Storage on-Demand class with support for the NFS and Lustre file systems, as well as the DAOS distributed object storage system. History of the “Govorun” supercomputer’s creation and development at JINR The “Govorun” supercomputer is the world’s first hyperconverged complex for high-performance computing with 100% hot-water liquid cooling, created by specialists from the RSC Group of Companies in 2018. Several stages of its modernization are completed. Building a hyperconverged system is based on pooling resources for computing and storage at each node of the system. This allows optimizing the scaling of the system’s total resources while increasing in the number of nodes. The operation of the “Govorun” supercomputer’s first stage made it possible to perform a number of complex resource-intensive calculations in the field of lattice quantum chromodynamics to study the properties of hadronic matter at high energy density and baryon charge and in the presence of supramaximal electromagnetic fields, to qualitatively enhance the efficiency of modeling the dynamics of collisions of relativistic heavy ions, to speed up the process of event generation and reconstruction for conducting the experiments within the NICA megaproject, to carry out computations of the radiation safety of experimental facilities, to significantly accelerate radiation biology studies and other applied tasks solved at JINR. The results of the studies are published in over 50 world’s leading scientific journals. In 2019, the JINR supercomputer was modernized. The updated computing complex had a total theoretical peak performance of 860 TFlops of double precision, while being a hyperconverged software-defined system. This approach is novel to the high-performance systems industry, and the project implemented using it has unique properties in terms of flexibility in customizing the system for a specific user task and, thus, allows for a significant increase in the efficiency of utilizing the supercomputer’s resources. In 2022, during the next planned modernization, the performance of the “Govorun” computing system at JINR reached 1.1 PFlops (petaflops is a quadrillion or 1015 floating-point operations per second). The application of liquid cooling in the JINR supercomputer enables not only to significantly save electricity during the functioning of the computing complex (less than 3% of the electricity required for the operation of the entire system is spent on its cooling), but also to employ the most high-performance server processors, Intel Xeon Platinum 8368Q, the suffix Q means that they are specially designed for liquid cooling and are superior to other models oriented towards the use of previously widespread air cooling. The use of a hyperconvergent approach made it possible to build for the “Govorun” supercomputer a unique high-speed data storage system with leading characteristics, namely, the speed of the parallel file system for reading/writing information exceeds 300 Gbps. Achieving such indicators became possible only due to the implementation of complex software and hardware solutions from the RSC, since the application of standard technologies for building stand-alone data storage systems would entail a solution tens of times more expensive. The unique implemented solution of the on-demand storage system allows the “Govorun” supercomputer to form a specific storage system for each task launched on the supercomputer with required properties, such as volume and speed, file system type, lifetime, reliability and security level, which would also be impossible in the case of a standard approach to building a data storage system for high-performance computing systems.. About the RSC Group of Companies The RSC Group of Companies is a leading Russian developer and integrator of full-cycle innovative ultra-dense, scalable, energy-efficient hyperconverged solutions for supercomputers and high-performance computing (HPC), data centers, intelligent on-demand data storage systems on top of various processor architectures and advanced liquid cooling, along with the Group’s many innovations. Since 2018, the RSC has been a participant in the National Champions priority project, implemented in the Russian Federation. The company’s existing potential enables to create the most energy-efficient solutions with record power usage effectiveness (PUE), to implement the highest computing density in the industry on top of standard x86 architecture processors and domestic Elbrus-16S, to use a completely “green” design, to ensure the highest reliability of the solution, the complete noiselessness of computing modules, 100% compatibility and guaranteed scalability, while achieving an unprecedentedly low cost of ownership and low energy consumption. In addition, RSC specialists have experience in developing and implementing an integrated software stack of solutions to enhance the efficiency of operation and applied use of supercomputer complexes, from system software to vertically-oriented platforms based on cloud computing technology. Additional information is available at the corporate website www.rscgroup.ru. RSC, RSC Exastrim, RSC BasIS, RSC Tornado, PetaStream, ScaleStream, and the RSC logos are registered trademarks of the RSC Group of Companies in Russia, the USA, Japan, and many European countries.
