Core Viewpoint - The article emphasizes the significant growth potential in China's switch chip market driven by increased AI spending and the trend towards Scale up technology, with projected market sizes of 257 billion, 356 billion, and 475 billion yuan for the years 2025, 2026, and 2027 respectively, reflecting year-on-year growth rates of 61%, 39%, and 33% [1] Group 1: Market Dynamics - The overall domestic localization rate of switch chips is currently low, particularly in the high-end chip market, where companies like Broadcom, Marvell, and NVIDIA dominate, indicating a substantial space for domestic manufacturers to replace imports [1] - The continuous evolution of large models and the expansion of Scale up clusters are identified as key trends, with large language model (LLM) parameters evolving from hundreds of billions to trillions and beyond, necessitating advanced strategies to address the challenges of model size [1] Group 2: Technological Developments - The article notes that the Scale up network bandwidth significantly exceeds that of Scale out networks, making it the mainstream technical solution for the industry, particularly in addressing the high bandwidth and low latency requirements of tensor and expert parallelism [1] - Overseas AI chips are upgrading their Scale up capabilities, with current GPU interconnects reaching dozens of cards and evolving towards hundreds, while AI custom chips are moving from dozens to thousands of cards, leading to increased demand for Scale up switch chips to enhance communication efficiency [2] Group 3: Domestic Innovations - Domestic AI companies have launched their own super node products equipped with Scale up switch nodes, with notable examples including Huawei's Ascend achieving interconnects of 384 chips and Baidu's Kunlun supporting 32/64 card interconnects [2] - Companies like ZTE, H3C, and Super Fusion have also introduced their super node solutions, providing foundational engineering capabilities for domestic chips to transition to super nodes, with ZTE's super node server supporting GPU communication bandwidths ranging from 400GB/s to 1.6T/s [2] Group 4: Protocol Perspectives - In the Scale up switch domain, Ethernet, PCIe, and proprietary protocols (such as NVLink and UB) are expected to coexist, while in the Scale out domain, Ethernet is anticipated to dominate due to its open ecosystem and cost advantages, with InfiniBand capturing a portion of the market [2]
国泰海通|电子:Scale up带来交换芯片新场景,国产渗透空间广阔
国泰海通证券研究·2025-08-24 13:35