Specialists at the Russian Federal Nuclear Center VNIIEF, part of Rosatom, have developed a high-speed communication system for supercomputers, data centers, and AI platforms. The system achieves a nominal data transfer rate of 200 Gbit/s, with peak throughput reaching 300 Gbit/s.
The solution is built on a proprietary domestic microchip. It includes a PCI Express 3.0 x16 controller, high-speed transceivers with up to 25 Gbit/s per lane, and a switch. The architecture supports GPU Direct technology, enabling direct data exchange between GPUs and memory without involving the CPU.
The system also features remote direct memory access to reduce latency and CPU load. These capabilities are designed to improve performance in distributed neural network training and large language model development, where inter-accelerator communication is critical.
The technology is intended for use in supercomputer clusters, cloud platforms, parallel data storage systems, and AI infrastructure. Potential applications include scientific calculations, climate modeling, rendering, and financial analytics.
The new system will be showcased at the Microelectronics forum in Sirius, held from September 27 to October 3, 2026. VNIIEF plans to continue expanding the system's functionality and adapting its architecture for future high-performance computing tasks.
