计算机：端侧大模型近存计算，定制化存储研究框架

Investment Rating - The industry investment rating is "Strong Buy" which is maintained [1] Core Insights - The large language models (LLMs) empower on-device AI, with a global market expected to grow from $15.2 billion in 2022 to $143.6 billion by 2032, indicating a nearly tenfold increase [4][19] - The maturity of compute-in-memory (CIM) technology provides a technical foundation for the commercialization of on-device AI large models [4] - The NPU (Neural Processing Unit) enhances the capabilities of on-device large models, addressing the diverse computational needs of generative AI applications [4] - Heterogeneous computing architecture requires advanced packaging technology to support high-performance, energy-efficient, and multifunctional computing systems [4] Summary by Sections On-Device Large Models - The feasibility and importance of running LLMs on edge devices have been validated with the emergence of models with fewer than 10 billion parameters, such as Meta's LLaMA and Microsoft's Phi series [12] - New models are increasingly being launched, enhancing capabilities in text processing and multimodal inputs [12] Near-Memory Computing - Near-memory computing (PNM) integrates storage and computation to enhance efficiency, with applications in AI, big data, and edge computing [26] - PNM can be categorized into two approaches: moving storage closer to processors and offloading data processing to storage [28] DRAM Technology Development Path - The report discusses various DRAM technology routes, including SRAM, DRAM, and Flash, highlighting their advantages and limitations in the context of compute-in-memory solutions [27] Advanced Packaging - Advanced packaging technologies are essential for integrating different functional chips to achieve high performance and energy efficiency [4] Customized Storage: Winbond CUBE Introduction - The CUBE technology developed by Winbond is a customized high-bandwidth memory solution designed for edge AI computing, offering low power consumption and compact size [5] - CUBE architecture allows for efficient integration of SoC and DRAM, reducing costs and enhancing performance [5] Related Companies - Recommended companies to watch include: - Storage: GigaDevice - Digital: Rockchip, Cambricon, Guokai Micro, Beijing Junzheng, Allwinner Technology, and Juchip Technology - IP: Chipone - Packaging: JCET, Tongfu Microelectronics, Huatian Technology, Nexperia, and Jingfang Technology [5]