Workflow
AI芯片Scale up
icon
Search documents
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
Core Insights - The article emphasizes the growing demand for switch chips driven by the expansion of AI chip Scale up clusters, with significant opportunities for domestic manufacturers in the high-speed chip market [1][2]. Group 1: Market Trends - The total scale of China's switch chips is projected to reach 257 billion, 356 billion, and 475 billion yuan in 2025, 2026, and 2027 respectively, with year-on-year growth rates of 61%, 39%, and 33% [1]. - The current low domestic production rate of switch chips, especially in the high-end market, presents a substantial opportunity for domestic manufacturers to increase market share as they continue to make breakthroughs in high-speed rates [1]. Group 2: Technological Developments - The evolution of large language models (LLMs) is leading to the development of various strategies such as data parallelism and tensor parallelism to address the challenges of model size, necessitating high bandwidth and low latency networks [1]. - The Scale up network's bandwidth significantly exceeds that of Scale out networks, making it the preferred technical solution in the industry [1]. Group 3: Domestic Innovations - Domestic AI companies are launching their own super-node products equipped with Scale up switch nodes to enhance communication efficiency [2]. - Companies like Huawei, Baidu, and ZTE are developing super-node solutions that support high interconnectivity, with ZTE's super-node server achieving GPU communication bandwidth of 400GB/s to 1.6T/s [2]. Group 4: Protocol Landscape - In the Scale up switch domain, Ethernet, PCIe, and private protocols (such as NVLink and UB) are expected to coexist, while Ethernet is likely to dominate the Scale out domain due to its open ecosystem and cost advantages [3].