半导体行业跟踪报告之二十四：络交换是AI集群互联核心，盛科通信交换芯片国内领先

Investment Rating - The report maintains a "Buy" rating for the electronic industry [1] Core Insights - The AI model training is increasingly reliant on high-performance networking technologies such as InfiniBand and Ethernet, which are essential for connecting AI clusters [2][3] - The rapid expansion of model parameters in AI necessitates significant improvements in computational power and memory, leading to the adoption of distributed training techniques [2][3] - The report highlights the competitive landscape, with major players like NVIDIA and Broadcom actively developing technologies and forming alliances to enhance GPU interconnect standards [21][25] Summary by Sections Section 1: AI Model Training and Networking - InfiniBand and Ethernet RoCEv2 are identified as the primary interconnect technologies for AI model training clusters, addressing the challenges of latency, bandwidth, and scalability [3][4] - The report emphasizes the need for high-performance networks to ensure efficient communication between computing nodes in AI clusters [2][3] Section 2: Market Position of Key Players - NVIDIA leads the InfiniBand market, with significant advancements in network card and switch technologies, including the deployment of 400Gbps switches [4][6] - The Ethernet switch chip market is dominated by Broadcom, Marvell, and other key players, with a high concentration of market share [12][13] Section 3: Domestic Leader - 盛科通信 (Centec) - 盛科通信 is positioned as a leading domestic Ethernet switch chip designer, breaking the foreign monopoly and contributing to the development of AI cluster interconnect solutions in China [26] - The company has established a comprehensive product line of Ethernet switch chips, catering to various network applications and achieving significant market presence [26][29] Section 4: UALink Alliance - The formation of the UALink alliance by major tech companies aims to create open standards for GPU interconnects, challenging NVIDIA's market dominance [21][22] - The alliance's first version of UALink is expected to enhance data transfer efficiency between different GPU architectures, promoting collaboration and innovation in the industry [22][25] Section 5: Financial Performance of 盛科通信 - In the first three quarters of 2024, 盛科通信 reported revenues of 808 million yuan, a year-on-year decrease of 7.95%, with a net loss of 76 million yuan [29][36] - The company continues to focus on high-performance Ethernet switch chips, with ongoing development of advanced products to meet the growing demands of data centers and enterprise networks [36][50]