公众号记得加星标⭐️，第一时间看推送不会错过。 过去几年，云厂商为了训练大模型投入巨资购买芯片，如今也到了利用推理实现变现的时候了。根据麦肯锡报告，全球AI推理市场规模预计2028年将达1500亿美 元，年复合增长率超40%，远高于训练市场的20%。推理支撑着各类应用的实时推理需求，包括智能推荐、内容生成、虚拟助手等。可以说，推理阶段才是实现实 际应用和商业化的关键。 这场推理之战，随着华为、英伟达和谷歌三大巨头相继发布了各自的推理芯片之后，已经将正式打响！ 华为Ascend 950PR： 成本优化下的推理利器 9月18日，在2025年华为全联接大会上，华为宣布了昇腾芯片的规划和进展。未来3年，也就是到2028年，华为在开发和规划了三个系列，分别是Ascend 950系列、 Ascend 960、Ascend 970系列。华为表示，将以几乎一年一代算力翻倍的速度，同时围绕更易用，更多数据格式、更高带宽等方向持续演进，持续满足AI算力不断 增长的需求 以往每年9月，都是手机发烧友的狂欢月，因为这时期苹果、小米、华为等都会发新机。然而，今年的9月，一个更深层次的产业变革正在暗流涌动。当所有人都在 对iphone ...

Core Insights - Amkor Technology announced a significant investment to build a state-of-the-art semiconductor advanced packaging and testing campus in Arizona, increasing total investment to over $7 billion [2][3] - The project aims to enhance U.S. semiconductor leadership and will create up to 3,000 high-quality jobs [2][3] Investment Details - The expansion includes new cleanroom space and a second packaging factory, with a total investment exceeding $5 billion, bringing the overall investment to $7 billion [2] - The project will be implemented in two phases, with the first production facility expected to be completed by mid-2027 and operational by early 2028 [2] Strategic Importance - The new campus will focus on advanced packaging and testing technologies, complementing TSMC's front-end wafer manufacturing, thus establishing a complete domestic semiconductor manufacturing ecosystem [3] - The facility is strategically located in Arizona's high-tech corridor, leveraging local talent, infrastructure, and industry clusters [3] Industry Support - The project has received support from the CHIPS for America Program and advanced manufacturing investment tax credits [2] - Key industry leaders, including executives from Apple and NVIDIA, expressed their support, highlighting the importance of this investment for the U.S. semiconductor supply chain and innovation [3]

公众号记得加星标⭐️，第一时间看推送不会错过。 （lateral p-GaN HEMT）基准技术平台。目前，围绕 p 型氮化镓刻蚀和欧姆接触形成的工艺模块研发工作正在推进中。后续，项目将进一步瞄准高压应用场景：针 对 650 伏特及以上的高压需求，研发工作将采用 300 毫米半规格衬底，以及与 CMOS 兼容的 QST工程衬底（一种以多晶氮化铝（AlN）为核心的材料）。在整个研 发过程中，300 毫米晶圆的翘曲度控制及其机械强度是需要重点关注的核心问题。 imec 在成功完成 300 毫米晶圆处理测试与掩模版研发后，正式启动了此次 300 毫米氮化镓项目，并计划于 2025 年底前在其 300 毫米洁净室中全面部署相关技术能 力。斯特凡・德库泰尔补充道："300 毫米氮化镓技术研发的成功，还取决于能否构建一个稳健的生态系统，并联合推动从 300 毫米氮化镓生长、工艺集成到封装解 决方案的全链条创新。因此，我们非常荣幸地宣布 AIXTRON、格芯、科磊、新思科技和维易科成为我们 300 毫米氮化镓开放式研发项目的首批合作伙伴，也期待 未来能有更多伙伴加入。因为先进氮化镓电力电子器件的研发，需要设计、外延、工艺集 ...

Core Insights - The article discusses the growing importance of Gallium Nitride (GaN) and Silicon Carbide (SiC) as third-generation semiconductor materials, particularly in the context of AI data centers, where they are creating new market opportunities [1][30]. - The rise of AI is significantly increasing power demands in data centers, necessitating upgrades in power supply systems to accommodate higher efficiency and power density [3][30]. Group 1: AI Data Center Power Challenges - The power consumption of AI data centers is projected to reach 7% of global energy consumption by 2030, equivalent to India's current energy usage [3]. - Traditional silicon-based devices have reached their performance limits, making wide bandgap semiconductors like SiC and GaN essential for meeting the demands of higher voltage, faster switching frequencies, and greater power density [3][30]. Group 2: Technical Advantages of SiC and GaN - SiC offers lower conduction resistance and stable temperature characteristics, making it suitable for high-voltage and high-temperature applications, particularly in AC-DC conversion [5]. - GaN achieves low switching losses and high switching frequencies, making it ideal for high-density applications in DC-DC conversion [5][30]. Group 3: Industry Leaders and Competitive Landscape - Infineon is positioned as a leader in power semiconductors, launching products like the CoolSiC™ MOSFET 400V series, which enhances power density and efficiency for AI server power supplies [7][8]. - Navitas Semiconductor combines SiC and GaN technologies to create high-power density solutions, recently introducing a 4.5kW AI data center server power solution with a power density of 137W/in³ and efficiency exceeding 97% [9]. - ON Semiconductor focuses on high output power, conversion efficiency, and power density, offering innovative solutions that balance small packaging with high performance [10]. Group 4: NVIDIA's Role in Driving Change - NVIDIA is seen as a key player in pushing the adoption of third-generation semiconductors, advocating for an 800V high-voltage direct current (HVDC) infrastructure in data centers [14][15]. - The shift to an 800V architecture is expected to create significant demand for new power devices and semiconductors, with NVIDIA's plans for future GPU and CPU deployments driving this transformation [15][16]. Group 5: Market Outlook - The market for GaN is expected to grow faster than SiC in AI data centers, driven by the demand for high-voltage applications and the advantages of GaN in high-frequency, low-loss scenarios [20][30]. - The article anticipates a golden era for third-generation semiconductors in AI data centers, contributing to technological advancements and more efficient infrastructure [30].

Core Insights - AMD and OpenAI have announced a strategic partnership involving the purchase and deployment of up to 6 gigawatts of AMD Instinct GPUs, potentially generating sales of up to $90 billion [1][2] - The partnership is seen as a structural shift in AI infrastructure financing, linking AMD's long-term valuation directly to OpenAI's growth [2][3] - AMD's stock surged over 30% following the announcement, indicating market confidence in the partnership's potential to accelerate AMD's AI revenue goals [2][16] Partnership Details - OpenAI will initially purchase 1 gigawatt of AMD Instinct MI450 GPUs, with the potential to scale up to 6 gigawatts over time [3][4] - The agreement includes a performance-based warrant allowing OpenAI to purchase up to 160 million shares of AMD at $0.01 each, contingent on achieving certain milestones [1][4] - The structure of the deal is designed to align incentives between AMD and OpenAI, rewarding AMD based on OpenAI's success in deploying the GPUs [4][10] Financial Implications - The partnership is expected to generate "hundreds of billions" in revenue for AMD, with the potential for significant stock price appreciation linked to performance milestones [2][9] - The warrant structure allows AMD to convert customer acquisition costs into equity, limiting dilution while ensuring a minimum shipment commitment of 1 gigawatt [5][10] - OpenAI's ability to monetize GPU capacity is crucial for the value of the warrants, creating a dependency on AMD's stock performance [10][13] Industry Context - The collaboration occurs in a competitive landscape dominated by NVIDIA, with OpenAI's existing partnerships with Microsoft and Oracle further complicating the dynamics [11][12] - AMD's approach contrasts with NVIDIA's and Intel's models by embedding equity incentives rather than requiring direct capital investment, fostering a more distributed network of strategic buyers [11][12] - The partnership may attract smaller AI developers seeking long-term relationships with semiconductor partners without the capital burdens of larger ecosystems [12] Technical Roadmap - The MI450 GPU, based on the CDNA4 architecture, is positioned as AMD's first product optimized for rack-level deployment, competing directly with NVIDIA's offerings [7][8] - AMD's Helios platform integrates MI450 GPUs with Zen 6 CPUs, aiming to provide a comprehensive solution for large AI clusters [7][8] - Future developments may include the MI500 series, which will build on the same design principles and enhance memory bandwidth and integration [8][9] Market Reaction - Following the announcement, AMD's stock price increased significantly, reflecting investor optimism about the partnership's implications for AMD's growth trajectory [16][17] - Analysts view the deal as a validation of AMD's AI strategy, positioning its data center division as a comparable growth engine to NVIDIA's [16][17] - The market response suggests confidence in AMD's ability to capture high-value AI workloads through its upcoming GPU architectures [16][17] Strategic Considerations - The partnership's success hinges on OpenAI's ability to scale its operations and effectively utilize AMD's hardware, which may influence AMD's broader market strategy [20] - AMD's recent acquisitions aim to strengthen its position in the AI value chain, enhancing its capabilities from chips to systems and software [18][19] - The collaboration with OpenAI could accelerate the evolution of AMD's ROCm ecosystem, providing valuable feedback for improvements [19][20]

公众号记得加星标⭐️，第一时间看推送不会错过。 NAND Flash 也呈现类似趋势。2026 年 NAND Flash 需求增长率预计为 13.8%，生产增长率预计为 14.0%，供需将保持平衡。 iM 证券分析师宋明燮指出："HBM 在 2024-2025 年经历了爆发式增长，但 2026 年随着竞争对手进入市场、供应规模扩大，其增长速度很可能放缓。" 与之相对，他 分析称："长期处于价格下跌态势的 NAND Flash，目前已触底，未来有相对走强的空间。" 智能手机、服务器需求成关键变量 需求层面，智能手机与服务器被视为核心影响变量。 来源 ： 内容来自 编译自zdnet 。 过去几年拉动行业景气度的高带宽内存（HBM），将因竞争加剧而增长放缓；反观动态随机存取内存（DRAM）与闪存（NAND Flash），则有望凭借供需平衡实 现良好业绩。 根据 iM 证券 6 日发布的报告，2025 年 DRAM 需求增长率已上调至 19.3%，小幅高于行业 18.1% 的生产增长率。展望 2026 年，受全球经济放缓影响，DRAM 需求 增长率预计为 14.1%，生产增长率预计为 14.2%。 https://z ...

公众号记得加星标⭐️，第一时间看推送不会错过。 来源 ： 内容编译自tomshardware 。 75 年前，贝尔实验室（Bell Labs）发明晶体管的三位科学家终于拿到了他们的美国专利。这个看起 来微不足道的三电极半导体器件点燃了第三次工业革命。更重要的是，它开创了硅与软件的时代，这 一时代至今仍在主导商业和人类社会。 （图片来源：贝尔实验室） 第一个可工作的晶体管于 1947 年问世，但直到 1950 年 10 月 3 日，约翰·巴丁（John Bardeen）、 沃尔特·布拉顿（Walter Brattain）和威廉·肖克利（William Shockley）才获得了专利。该专利是针 对一个"利用半导体材料的三电极电路元件"颁发的。晶体管对商业和社会产生的重大影响，又过了几 年才得以充分实现。 晶体管取代了笨重、易碎且耗电的电子管。尽管电子管在某些吉他放大器、发烧级音响系统和录音棚 设备中依然存在，因为人们有时更喜欢它们"有机"的声音特性。我们也在一些军事、科学、微波/射 频应用中仍能看到电子管，因为在这些场景下，晶体管可能容易受到辐射或其他干扰。当然，还有其 他小众用途。 图片来源：英特尔 显然，在 ...

供应链牵引成为行业共识 今年3.28日昂瑞微科创板IPO申报获得受理，这在射频前端行业是个重大事件， 经过最近几年发 展，昂瑞微业绩发展迅速，2024年销售额赶上唯捷创芯，成功跻身射频前端第一梯队，与唯捷创 芯一道率先推出射频前端中最难的Sub3G L-PAMiD产品，并在头部品牌手机终端客户旗舰机型量 产，填补射频前端最关键的一块短板。据昂瑞微披露资料，其在卫星通信和车载电子等应用的射频 前端市场也有较大进展，5G接收模组也在头部品牌手机终端量产，补齐5G射频前端拼图。 美国打压成就中国射频前端公司迅速发展 昂瑞微近几年发展迅速，离不开美国对中国通信巨头华为、中兴打压带来的第一波国产替代红利。 在2019年之前，国产射频厂家的日子不好过，做出来的芯片得不到大客户的认可，只能在中低端 市场厮杀，没有足够的财务实力并且没有头部品牌客户牵引，产品往往落后海外友商一代以上，竞 争力严重不足。这个情况在2019年以后得到改观，美国的打压让国内头部手机终端客户认识到对 国产供应链的牵引重要性，昂瑞微正是抓住这波机会，紧跟头部大客户牵引，在2023年成功开发 出可以用于高端旗舰手机的射频前端芯片系列，完成华丽转身，跻身国 ...

Core Viewpoint - Samsung's CMOS image sensor (CIS) will return to Apple's supply chain, marking the first time in a decade that Samsung has been involved in this capacity, with potential demand estimated at around 200 million units [3][4]. Group 1: Product Details - The flagship iPhone will feature one front CIS and three rear CIS, with one of the rear lenses supplied by Samsung [3]. - Current orders from Apple to Samsung only account for about 20% of the maximum demand, based on an 80% yield from 12-inch wafers [3]. Group 2: Production Timeline - Samsung originally planned to start mass production in March 2026, but this has been pushed back to late 2026 or early 2027 [3]. - Due to the delay, the impact on Samsung's revenue from the iPhone 18 series will not be immediate, with the iPhone 18 models expected to launch in the second half of 2026 [3]. Group 3: Competitive Landscape - Samsung is developing a 50 million pixel CIS for Apple, focusing on stability over the introduction of new technology [4]. - Although Samsung is perceived to lag behind Sony in low-light performance, it has a competitive edge in pixel technology, having achieved 200 million pixel resolution through its ISOCELL brand in 2021 [4]. - There is optimism that Samsung could secure a unique supply role for high-pixel CIS for iPhones and potentially become a second supplier for the 50 million pixel CIS product line traditionally dominated by Sony [4].

Core Viewpoint - The Bay Area Semiconductor Industry Ecological Expo ("Bay Chip Expo") will take place from October 15-17, 2025, in Shenzhen, showcasing over 600 leading global companies and aiming to tap into a billion-level market in the Bay Area through a strategy of "full industry chain linkage and precise matching" [1][3]. Group 1: Event Overview - The Bay Chip Expo will cover an exhibition area of 60,000 square meters and feature a joint exhibition area of 60 square meters, highlighting four key companies: Zhihe Computing, Tektronix, Yuntian Lifeng, and Ouyue Semiconductor [1][3]. - The expo will focus on "AI chip ecology" and "advanced packaging," covering the entire core link from architecture innovation to application in automotive regulations [3]. Group 2: Company Highlights - **Zhihe Computing**: Focuses on developing high-performance, scalable "integrated CPU" chips based on RISC-V architecture for AI applications, with its first product, the "Archimedes" series, designed for efficient general computing and AI inference [3][5]. - **Tektronix**: Known for providing precise and straightforward testing solutions, Tektronix will demonstrate a full-process solution from wafer-level material testing to chip design verification at the expo [4][8]. - **Yuntian Lifeng**: A leader in AI inference chips, the company has developed neural network processor core IP and chips, focusing on efficient inference for large models and enabling AI applications across various sectors [5][8]. - **Ouyue Semiconductor**: The first Chinese company to focus on intelligent automotive third-generation E/E architecture, Ouyue aims to enhance smart technology in practical scenarios, offering a range of products for intelligent automotive components [6][8]. Group 3: Interactive Opportunities - The expo will feature three interactive segments: daily technical salons at 10:00 AM, targeted supply-demand matching meetings at 2:00 PM, and customized technical solution consultations throughout the event [7][8]. - The event will provide opportunities for attendees to engage with experts and core buyers from companies like Huawei and BYD, facilitating direct connections and potential collaborations [7][8].