半导体：AI需求推动运力持续增长，互联方案重要性显著提升

Investment Rating - The report maintains an "Outperform" rating for the semiconductor industry [1]. Core Insights - The demand for AI is driving continuous growth in computing power, highlighting the importance of interconnection solutions [1]. - Rapid development of AI applications is accelerating the demand for "computing power" and "storage capacity," while also raising the requirements for "interconnectivity" [1]. - The construction of interconnect communication capabilities for AI computing clusters includes three main aspects: Die-to-Die, Chip-to-Chip, and Board-to-Board interconnections [20][21]. Summary by Sections Interconnection Technology and Chip Demand - The interconnection technology is evolving rapidly, stimulating demand for interconnect chips [1]. - PCIe Retimer chips are becoming mainstream solutions in the industry, with a typical AI server configuration requiring 8 to 16 PCIe Retimer chips [1][46]. - CXL MXC chips are core components for high-bandwidth memory expansion modules, significantly increasing memory capacity and bandwidth for data-intensive applications [1][63]. - MRCD/MDB chips are essential for meeting the high bandwidth memory requirements of AI and big data applications [1][92]. Market Opportunities - Companies such as 澜起科技 (Lianqi Technology), 盛科通信 (Shengke Communication), and 裕太微 (Yutai Micro) are recommended for attention within the interconnect industry chain [1]. - The global CXL market is expected to reach $15 billion by 2028, driven by the increasing compatibility of CPUs with CXL standards [85][86]. Performance Enhancements - The report discusses two main methods for enhancing data center performance: Scale-up (vertical scaling) and Scale-out (horizontal scaling) [24]. - The introduction of NVLink technology has addressed bandwidth and transmission bottlenecks associated with PCIe, facilitating high-speed connections between multiple GPUs [28][13]. Future Trends - The report anticipates that the market for PCIe Retimer chips will continue to expand as GPU demand increases [1][109]. - The integration of CXL technology is expected to enhance system performance and flexibility, with predictions indicating that by 2027, all CPUs will support CXL interfaces [86].