公众号记得加星标⭐️，第一时间看推送不会错过。 英特尔本周宣布，正与软银旗下子公司SAIMEMORY合作，将Z角内存（ZAM）商业化。ZAM是一 种先进的DRAM，可将内存模块垂直堆叠。虽然ZAM芯片预计至少三年内无法上市，但最终可能会 取代目前因人工智能蓬勃发展而需求旺盛的高带宽内存（HBM）。 内存带宽目前是人工智能处理的主要瓶颈，因为各组织机构都在寻求将越来越多的数据从内存传输到 GPU，然后再传输回内存。英伟达和AMD等芯片制造商正在其GPU芯片上集成数百GB的HBM内 存，以缓解这一瓶颈。然而，HBM需求的激增导致全球NAND闪存库存短缺，推高了内存模块和 NVMe存储的价格，并造成了供应链短缺。 NGDB 计 划 是 先 进 存 储 技 术 (AMT) 项 目 的 一 部 分 ， 该 项 目 是 美 国 能 源 部 和 国 家 核 安 全 管 理 局 (NNSA) 的一项倡议，旨在将英特尔、SK 海力士和软银等供应商与能源部政府实验室聚集在一起， 开发新的存储技术，包括 ZAM、HBM、Compute Express Link (CXL) 和非易失性存储器，如磁性 随机存取存储器 (MRAM)。 ...

在AI算力赛道，英伟达凭借Hopper、Blackwell、Rubin等架构GPU，早已在AI训练领域建立起了难以撼动的技术壁垒与行业地位。但随着 即时AI场景需求爆发，传统GPU在面对低批处理、高频交互推理任务中的延迟短板愈发凸显。 为破解这一痛点，英伟达重磅出击，斥资200亿美元收购Groq核心技术，抢跑AI推理市场。 这一金额不仅创下英伟达历史最大手笔交易、刷新了推理芯片领域的估值纪录，更鲜明地昭示着英伟达从"算力霸主"向"推理之王"转型的意志。 紧随这一动作，据技术博主AGF消息进一步披露，英伟达计划在2028年推出新一代Feynman架构GPU——采用台积电A16先进制程与SoIC 3D堆叠 技术，核心目的正是为了在GPU内部深度集成Groq那套专为推理加速而生的LPU（语言处理单元），相当于给GPU加装了一个专门处理语言类推理 任务的专属引擎，直指AI推理性能中长期存在的"带宽墙"与"延迟瓶颈"。 回看中国市场，AI浪潮推动下，国产大模型多点突破、强势崛起，本土AI芯片企业集体爆发并密集冲击IPO，资本热度居高不下。 然而，当英伟达选择通过Feynman架构来补齐推理短板时，就意味着谁能率先解决" ...

Core Insights - The article discusses the shift in memory technology for Microcontroller Units (MCUs) as embedded flash memory (eFlash) reaches its limits at 28nm, prompting manufacturers to explore new storage types like MRAM, PCM, RRAM, and FRAM [1][3][6]. Group 1: Industry Trends - The industry is moving towards new types of memory to enhance MCU performance, with MRAM being particularly favored due to its diverse types and broad application prospects [1][6]. - Major companies such as Huawei, TSMC, Samsung, Intel, and NXP are investing in MRAM technology, indicating strong industry interest and potential growth [1][6][19]. Group 2: Technical Advantages of MRAM - MRAM offers a combination of speed, low power consumption, and high durability, making it suitable for various applications, including automotive and AI accelerators [10][15][20]. - The technology allows for word-level erase and program capabilities, providing an energy-efficient non-volatile memory solution [15][16]. Group 3: Product Developments - Infineon has launched the AURIX TC4x series MCU using RRAM technology, while STMicroelectronics has introduced the xMemory Stellar series MCU with PCM [5][6]. - NXP's S32K5 MCU, the first 16nm FinFET+MRAM MCU, features high performance and low power consumption, integrating multiple ECUs into a single system [19][20]. - Renesas has released the RA8P1 series MCU with MRAM, emphasizing high performance and durability compared to traditional flash memory [22][28]. Group 4: Future Outlook - The article suggests that MRAM's integration into MCUs is accelerating, with TSMC making strides in the industrialization of third-generation SOT-MRAM technology [33]. - While MRAM presents significant advantages, it also faces challenges such as material complexity and sensitivity to strong magnetic fields, which may limit its application in certain environments [18][33].

Core Viewpoint - MRAM technology is emerging as a new generation of storage solutions, with Zhejiang Chituo Technology leading the development and manufacturing of MRAM chips in China, showcasing its latest products at the 2025 Shenzhen International Electronics Exhibition [1][9]. Group 1: MRAM Technology Advantages - MRAM utilizes magnetic materials to represent binary data, offering advantages such as high speed, low power consumption, high endurance, radiation resistance, and reliability [1]. - Chituo's MRAM products maintain data integrity for over ten years at 125°C, operate in a temperature range of -40 to +125°C, support over one trillion write cycles, and achieve a yield rate of 95% for large capacity arrays with sub-ppm failure rates [1]. Group 2: Embedded and Standalone MRAM - The embedded eMRAM can replace eFlash in MCU/SoC applications, with industry consensus indicating that 28/22nm will be the last cost-effective nodes for eFlash, while eMRAM can extend to 28nm and beyond [2]. - Chituo's eMRAM combines DRAM-like read/write speeds, non-volatility of flash memory, and SRAM-compatible interface characteristics, making it suitable for high-performance applications in industrial control, automotive electronics, identity authentication, and smart wearables [2]. Group 3: Product Series and Applications - Chituo's standalone MRAM is categorized into multiple series based on capacity, interface, and packaging, and has been adopted by leading users in various industries such as industrial control, power, and metering [5]. - The company is also at the forefront of research on the next generation of MRAM, specifically Spin-Orbit Torque MRAM (SOT-MRAM), and has proposed a groundbreaking device structure suitable for large-scale manufacturing [7]. Group 4: Company Overview - Zhejiang Chituo Technology is the first company in China to achieve mass production of MRAM, with a 12-inch MRAM pilot production line and a comprehensive platform for the research and industrialization of new storage chips [8].

Core Viewpoint - The semiconductor industry is shifting from a singular focus on process miniaturization to diversified innovation, with advanced packaging technologies and specialty processes driving performance optimization and differentiation in the market [1][2]. Group 1: Market Trends and Growth - The global specialty process market has surpassed $50 billion, with a compound annual growth rate (CAGR) of 15%, significantly outpacing the average growth rate of the semiconductor industry [1]. - Companies like TSMC, UMC, and SMIC are accelerating their investments in specialty processes, with TSMC establishing itself as a global benchmark through its extensive technology portfolio [2][4]. Group 2: TSMC's Specialty Process Landscape - TSMC offers a comprehensive range of specialty processes, including automotive, ultra-low power (ULP)/IoT, RF, embedded non-volatile memory (eNVM), high-voltage display, and CMOS image sensors (CIS) [4]. - TSMC's automotive-grade processes are designed for high reliability and long lifecycle, supporting advanced driver-assistance systems (ADAS) and smart cockpit applications [4]. - The N4e process is optimized for ultra-low power IoT AI devices, balancing performance and cost effectively [4]. Group 3: Innovations in Non-Volatile Memory (NVM) - TSMC is addressing the limitations of traditional eFlash technology by advancing embedded NVM technologies such as RRAM and MRAM, which are expected to replace eFlash in automotive and IoT applications [6][7]. - RRAM technology is being commercialized, with TSMC's 22nm RRAM already certified for automotive applications, and 12nm RRAM expected to follow suit [6][7]. - MRAM technology is also being developed for automotive applications, with NXP and TSMC collaborating on 16nm embedded MRAM for high-end automotive MCUs [20][21]. Group 4: Competitive Landscape and Future Directions - Major MCU manufacturers are exploring various new storage technologies, including eRRAM, eMRAM, ePCM, and eFeRAM, to enhance performance and reduce power consumption [16][31]. - The market for embedded NVM is projected to grow significantly, with wafer production expected to increase from approximately 3 KWPM in 2023 to about 110 KWPM by 2029, indicating a CAGR of around 80% [29]. - TSMC plans to integrate advanced processes with specialty technologies to support the evolution of chip architecture from "functional integration" to "system reconstruction" [8][34].

Core Viewpoint - The semiconductor industry is shifting from a singular focus on process miniaturization to diversified innovation, with advanced packaging and specialty processes becoming key drivers for performance optimization and differentiation [1][2]. Group 1: Specialty Processes and Market Growth - The global specialty process market has surpassed $50 billion, with a compound annual growth rate (CAGR) of 15%, significantly outpacing the average growth rate of the semiconductor industry [1]. - Specialty processes focus on customized and diverse process optimizations, achieving a precise balance of performance, power consumption, and cost, particularly in demanding fields like automotive electronics and IoT [1]. Group 2: TSMC's Leadership in Specialty Processes - TSMC is establishing itself as a global benchmark in specialty processes through a combination of technological breadth and ecosystem depth, expanding its capabilities across various domains including automotive and RF technologies [2][4]. - TSMC's advanced logic technologies, such as N7A, N5A, and N3A, are specifically designed for automotive applications, ensuring high reliability and long lifecycle [4]. Group 3: Innovations in Embedded Non-Volatile Memory (eNVM) - TSMC is addressing the limitations of traditional eFlash memory by advancing RRAM and MRAM technologies, which are expected to replace eFlash in automotive and IoT applications [6][8]. - The introduction of RRAM and MRAM technologies allows for significant improvements in performance, reliability, and power efficiency, with TSMC's RRAM already in mass production at 40, 28, and 22 nm nodes [7][8]. Group 4: Competitive Landscape and Future Trends - Major MCU manufacturers are collaborating with foundries to leverage specialty processes, with companies like Infineon and NXP adopting eNVM technologies to enhance their product offerings [9][16]. - The market for embedded NVM is projected to grow rapidly, with wafer production expected to increase from approximately 3 KWPM in 2023 to about 110 KWPM by 2029, indicating a strong shift towards new storage technologies [26]. Group 5: Diverse Storage Technologies - Various new storage technologies, including eRRAM, eMRAM, and ePCM, are being explored by different manufacturers, each offering unique advantages in terms of speed, power consumption, and integration capabilities [30][32]. - The trend indicates a move towards a multi-storage technology ecosystem rather than a single dominant solution, reshaping the MCU landscape in the post-eFlash era [32].