报告导读： 昇腾多模态推理加速套件正式开源，联合中科弘云发布 AI 推理加速联合解决 方案，加速昇腾芯片渗透率增长，昇腾链预计受益。 本订阅号所载内容仅面向国泰海通证券研究服务签约客户。因本资料暂时无法设置访问限制，根据《证 券期货投资者适当性管理办法》的要求，若您并非国泰海通证券研究服务签约客户，为保证服务质量、 控制投资风险，还请取消关注，请勿订阅、接收或使用本订阅号中的任何信息。我们对由此给您造成的 不便表示诚挚歉意，非常感谢您的理解与配合！如有任何疑问，敬请按照文末联系方式与我们联系。 法律声明 风险提示。 国产算力芯片需求增长不及预期；先进制程产能扩产不及预期。 报告来源 以上内容节选自国泰海通证券已发布的证券研究报告。 报告名称： 昇腾推理加速套件正式开源，昇腾芯片渗透加速；报告日期：2025.12.29 报告作者： 舒迪（分析师），登记编号：S0880521070002 段笑南（研究助理），登记编号：S0880124070028 重要提醒 投资建议。 根据 CNMO 科技， 2025 年 12 月 19 日华 为 昇 腾 多模态推理加速套件 MindIE SD 项目已正式开源，可有效提升推理效率 ...

Core Viewpoint - The price of High Bandwidth Memory (HBM) is expected to rise significantly, with SK Hynix confirming that the price for HBM4 will be approximately $560, over 50% higher than the current HBM3E price of around $370. This price increase raises concerns about dependency on high-end HBM, especially amid export controls affecting China. Huawei's newly open-sourced Unified Cache Manager (UCM) technology may provide a solution to mitigate this dependency by optimizing data management across different storage mediums [1][5]. Group 1: HBM Market Dynamics - SK Hynix leads the global HBM market with a 62% shipment share, followed by Micron Technology at 21% and Samsung Electronics at 17% [4]. - HBM4, the sixth generation of HBM, features a 2048-bit interface and up to 16 layers of stacking, targeting bandwidth exceeding 2 TB/s and capacity of 64GB [4]. - The integration of HBM directly with processors, including potential use of photonic technology, is being explored to enhance speed and efficiency [4]. Group 2: Huawei's Technological Innovations - Huawei's UCM technology allows for tiered caching of memory data based on usage frequency, which can optimize the efficiency of HBM and reduce costs [1][5]. - UCM architecture includes several key modules that enhance capabilities such as sparse attention and context window expansion, achieving up to 90% reduction in latency for the first token and a 22-fold increase in system throughput [1]. - Huawei has also developed its own HBM variants, HiBL 1.0 and HiZQ 2.0, which are designed to lower costs compared to high-performance HBM3e/4e, particularly for inference and recommendation scenarios [6].