Zhongjin: The construction of domestic AI computing infrastructure is accelerating, and the domestic high-speed interconnection network system is entering a breakthrough period.

Zhongjin released a research report stating that the AI computing power cluster is accelerating towards the scale of tens of thousands or even hundreds of thousands of cards. High-speed interconnection networks, as "neural hubs," have become a key link affecting the effective release of computing power in clusters. The current ecosystem of high-speed interconnection networks in the intelligent computing scene is mainly dominated by InfiniBand (IB) and RoCE v2, with core technologies and ecosystems being led by overseas manufacturers in the long term. With the accelerated development of domestic AI computing infrastructure, the domestic high-speed interconnection network system is entering a period of disruption, and domestically developed RDMA high-speed interconnection networks are expected to enter from the bottom-level architecture and self-developed hardware, promoting the gradual realization of domestic substitution in the interconnection aspect of domestic computing infrastructure. It is recommended to pay attention to the domestic computing industry chain. Zhongjin's main points are as follows: RDMA is an important technical path to achieve high-performance AI networks. The core feature of RDMA lies in "kernel bypass" and "zero copy," which can bypass the host's operating system kernel and CPU scheduling, directly access the memory area of the remote server for read and write operations, greatly reducing communication latency and reducing CPU resource consumption. Currently, mainstream RDMA solutions include IB, RoCE, and iWARP. IB is a native network designed for RDMA, with the best end-to-end lossless performance; RoCE applies the RDMA architecture to the Ethernet ecosystem, and RoCE v2 relies on prioritized traffic control and explicit congestion notification flow control mechanisms to optimize the network, simulating lossless transmission in traditional lossy Ethernet. Scale Fabric supports the localization of domestic supercomputing clusters with tens of thousands of cards. In the context of overseas manufacturers dominating the high-end interconnection network ecosystem, the domestic high-speed interconnection RDMA network is gradually breaking through, and the process of domestic substitution is accelerating. scale Fabric launched by Dawning Information Industry is a domestically developed 400G native lossless RDMA high-speed interconnection network architecture, utilizing credit flow control and link layer retransmission mechanisms consistent with InfiniBand. With native RDMA network cards and switching chips similar to IB, it achieves 400Gb/s ultra-high bandwidth, sub-1 microsecond end-to-end communication latency, and lossless transmission. Currently, scale Fabric as the network cornerstone for domestic supercomputing clusters with tens of thousands of cards has supported the landing of scaleX clusters with tens of thousands of cards, and is expected to help large-scale domestic computing clusters enter the accelerated deployment phase. In terms of investment recommendations, it is advised to pay attention to Dawning Information Industry (603019.SH). Risk factors include the maturity of the domestic computing interconnect ecosystem falling short of expectations and the development of domestic computing not meeting expectations.