美银：2026年及未来的人工智能互连-NVLink Fusion vs. UALink

Investment Rating - The report does not explicitly state an investment rating for the semiconductor industry or specific companies within it. Core Insights - The report highlights the competition between NVLink and UALink, with NVLink currently holding a significant advantage in AI scale-up interconnect technology. UALink is seen as a step forward for the non-NVDA ecosystem, aiming to catch up to NVLink's capabilities starting in 2027 [1][3][24]. - NVLink Fusion is noted to open up approximately $12 billion in total addressable market (TAM) for Nvidia in networking attached to custom accelerators, expanding its reach in the AI compute market [4][18]. Summary by Sections Industry Overview - UALink aims to compete with NVLink's dominance in AI scale-up interconnect technology, with both technologies enabling the connection of AI accelerators through fast, low-latency networking [1]. - UALink is generally 1-2 generations behind NVLink but represents a significant advancement for the non-NVDA ecosystem [1][3]. UALink Overview - UALink utilizes AMD's Infinity Fabric technology as its standard protocol, allowing for shared memory coherence between CPUs and GPUs [2][26]. - The UALink Consortium, founded in May 2024, oversees the development of UALink technology, with founding members including major players like AMD, Broadcom, and Intel [25]. NVLink vs. UALink Comparison - UALink 1.0 supports up to 800 GB/s bandwidth per accelerator, while NVLink 5.0 supports 1.8 TB/s, indicating that NVLink maintains a significant performance edge [3][31]. - By the time UALink-compatible hardware is expected to launch in 2026-27, NVLink 6.0 is anticipated to offer even greater bandwidth capabilities, further extending NVDA's lead [33][34].