半导体行业观察

Search documents
马斯克:AMD相当不错,但是……
半导体行业观察· 2025-09-13 02:48
公众号记得加星标⭐️,第一时间看推送不会错过。 来源 : 内容来自wccftech 。 亿万富翁埃隆·马斯克近日在推文中评价了 AMD 的 AI 硬件性能,称其足以支撑小型和中型 AI 模型 的运行,但在高端算力任务上,NVIDIA 依然占据优势。 AMD 的崛起与马斯克的使用场景 据悉,马斯克旗下的 xAI 已经使用了 AMD 的 Instinct MI300/MI300X AI 加速卡来支撑其 AI 模 型。此外,特斯拉也曾与 AMD 有过合作。但总体而言,NVIDIA 依旧是马斯克在 AI 需求上的核心 合作伙伴。不过,在小到中等规模的模型应用中,AMD 的 AI 性能已获得马斯克的肯定。虽然在大 规模模型训练方面,AMD 并非首选,但在推理、微调以及中型基础模型等场景,AMD Instinct 系 列已逐渐在行业内获得应用。 AMD 与大科技公司的差距 AMD 尚未能像 NVIDIA 一样强势打入微软、Meta 或谷歌等大型科技公司的 AI 供应链。虽然这些 公司在一定程度上也会采用 AMD Instinct GPU,但远不及 NVIDIA 的普及度。这意味着 AMD 仍 需付出更多努力来争取关注,而 ...
国巨,又收购了
半导体行业观察· 2025-09-12 01:14
国巨集团董事长陈泰铭再出手,11日召开董事会通过以新台币每股229.8元公开收购茂达28.5%股 权,溢价幅度达20%,充分展现本次出价的诚意与吸引力,以国巨预定收购的最高股数计算,整体收 购规模上看48.9亿元。 国巨表示,本次公开收购案将向金管会申报并公告,公开收购期间拟自2025年09月12日起至2025年 10月01日止。预定对之最高收购数量为21,277,245股(约茂达已发行普通股股份总数之28.5%),最 低收购数量为3,733,000股(约茂达已发行普通股股份总数之5%)。本次公开收购成就条件为收购数 量达最低收购数量。 公众号记得加星标⭐️,第一时间看推送不会错过。 来源 : 内容来自 经济日报 。 电源管理IC方面,茂达在PC、绘图卡等应用的产品拉货力道强,同时还有电源供应器升级带来的相 关订单,另外,其储存应用的订单也表现不错。 *免责声明:本文由作者原创。文章内容系作者个人观点,半导体行业观察转载仅为了传达一种不同的观点,不代表半导体行业观察对该 观点赞同或支持,如果有任何异议,欢迎联系半导体行业观察。 茂达电子为功率集成电路(Power IC)设计公司,致力从事混合信号功率晶片与感测 ...
英伟达抢进SiC?什么情况
半导体行业观察· 2025-09-12 01:14
Core Viewpoint - NVIDIA is initiating a "power revolution" by increasing the output voltage of AI server chips from 54V to 800V, utilizing Silicon Carbide (SiC) components to enhance power efficiency and performance in server cabinets [1][2]. Group 1: Silicon Carbide (SiC) Overview - Silicon Carbide (SiC) is a compound semiconductor material made from silicon and carbon, offering a wider bandgap than silicon, high-temperature resistance over 200 degrees Celsius, and excellent heat dissipation, making it suitable for high-power applications [1]. - SiC has been used in various applications such as 5G communications, electric vehicles (EVs), charging stations, and renewable energy systems, enhancing energy efficiency and performance [1]. Group 2: NVIDIA's Power Revolution - NVIDIA's new generation processor plans to upgrade the silicon intermediary layer to SiC, aiming for full-scale production of the 800V high-voltage direct current (HVDC) data center architecture by 2027 [2]. - The power consumption of AI servers is expected to increase from kilowatt (KW) levels to 1 million watts (MW), indicating a nearly 100-fold rise in future power demand [2]. - The traditional 54V architecture is nearing its limits in terms of copper loss, space, and conversion efficiency, necessitating the shift to SiC for improved efficiency and power density [2]. Group 3: Challenges and Considerations - While SiC offers significant advantages for high-performance server cabinets, it does not completely replace silicon in chip design and manufacturing, as silicon remains essential for CPU and GPU designs [3]. - The adoption of SiC is hindered by its high cost and the previous low voltage requirements of data centers, which limited the realization of SiC's benefits [4]. - Implementing SiC components requires careful design and wiring, complicating the control and safety verification of power systems, and necessitating an upgrade of the entire supply chain to accommodate the new architecture [4]. Group 4: Industry Implications - SiC is becoming a key material upgrade for power supply manufacturers, who are integrating SiC power components into server power supply units (PSUs) and HVDC distribution modules, highlighting the importance of technical expertise in this transition [5].
英特尔将失去至强首席架构师
半导体行业观察· 2025-09-12 01:14
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CRN 。 随着 Ronak Singhal 即将离职,他将成为英特尔大约八个月内失去的第二位 Xeon 首席架构师,而英 特尔首席执行官 Lip-Bu Tan 正在推动变革以振兴其数据中心业务,包括任命一位新的业务领导人。 据 CRN 获悉,英特尔首席执行官陈立武 (Lip-Bu Tan) 正在推动变革以振兴数据中心业务,今年该 公司将失去第二位 Xeon CPU 首席架构师。 据知情人士透露,同事们最近获悉,英特尔高级研究员兼至强产品首席架构师罗纳克·辛加尔(Ronak Singhal) 将于本月底离开这家半导体巨头。由于未获得授权,他们要求不透露消息。 随着辛加尔即将离职,他将成为英特尔大约八个月内失去的第二位至强首席架构师。此前,同样担任 英特尔高级研究员的赛莱什·科塔帕利 (Sailesh Kottapalli)于今年 1 月离开公司,帮助领导高通重振 服务器 CPU 的工作。 本文发表后,英特尔向 CRN 证实了辛加尔将离职的消息,但拒绝进一步置评。辛加尔本人也拒绝置 评。 英特尔的一位前同事称辛加尔是"行业巨头","他为推动 ...
CIS,怎么走?
半导体行业观察· 2025-09-12 01:14
公众号记得加星标⭐️,第一时间看推送不会错过。 过去五十年来,像素的发展轨迹不断增加工艺复杂性,以实现成像所需的性能。在这一过程中,开发 和制造工艺工程师功不可没,他们成功实现了一系列针对图像传感器的工艺改进,以满足性能、良率 和成本方面的要求。这些改进包括使用新材料来减少串扰、增强光学性能并支持附加功能(图1)。 这些工艺可以沿用自其他产品(例如MIMcaps,图2),也可以采用新颖的结构,例如气隙背面栅极 (图3),应用于图像传感器。 产品开发过程伴随着一系列技术挑战,直至满足市场需求(例如,缩小像素尺寸以降低成本并增加阵 列尺寸),然后转向下一个市场需求(例如,增加高动态范围等功能)。逆向工程的目的是记录每个 制造商所采用的技术开发,并预测流程开发中即将出现的决策点。 堆叠技术就是一项赋能技术的例子。其发展轨迹是从前照式单金属CCD到多金属CMOS(用于增加功 能),再到背照式CMOS(用于改善光学响应),再到面对面堆叠CMOS(用于在限制芯片尺寸的同 时增加图像处理功能)。最后一种技术需要金属互连,最初是通过位于芯片边缘的硅通孔实现的,但 这正在被混合键合所取代,这种键合利用了材料的特性:两个抛光的S ...
HBM,另一场混战
半导体行业观察· 2025-09-12 01:14
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 chosun 。 SEMECS、韩华半导体和韩美半导体三家公司在高带宽存储器(HBM)核心半导体设备TC键合机 (TC Bonder)的竞争中处于白热化阶段,目前正加速混合键合机的量产准备工作,预计该设备将从 第六代HBM(HBM4)生产开始部分采用。三星电子和SK海力士已进行HBM4的量产测试,目前的 关键在于设备价格(比现有TC键合机高出两倍以上)以及与美国和荷兰制造的设备相比,能否确保 技术竞争力和稳定性。 据业内人士12日透露,三星电子子公司SEMECS正在与三星电子半导体(DS)部门的下一代HBM (高密度存储器)专职人员合作,对混合键合机设备进行评估。尽管目前尚未达到目标性能,但双方 正共同努力攻克DRAM堆叠和热控制等多项技术难题。SEMECS的目标是在今年年底或明年向三星电 子提供能够量产的混合键合机。 混合键合机被称为下一代HBM市场的"游戏规则改变者"。目前用于HBM制造的TC键合机通过对凸块 (例如焊料的导电凸块)施加热量和压力来堆叠芯片。相比之下,混合键合机无需单独的凸块即可连 接芯片,使其成为制造20层或更高层数 ...
晶圆厂支出预测:1.5 万亿美元
半导体行业观察· 2025-09-12 01:14
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 electronicsweekly 。 普华永道表示,2024 年至 2030 年期间晶圆厂的支出将超过 1.5 万亿美元,以满足 2030 年价值超过 1 万亿美元的市场需求。 这与过去 20 年的支出相当,主要得益于政府补贴和稳定供应链的努力。普华永道的报告主要将投资 激增归因于全球服务器和网络市场的增长。 人工智能工作负载正在推动全球服务器市场的增长,预计到 2030 年将达到 3000 亿美元以上。 报告还强调了供给侧设计和转型对于强化半导体生态系统的重要性,这得益于无晶圆厂公司、代工厂 和IP供应商之间的密切合作。 然而,该行业面临着日益扩大的人才缺口,预计到2030年,将需要额外10万名工程师来支持预期的 设计增长。 报告称,到2030年,服务器和网络领域预计将成为最大的半导体需求市场。 数据中心人工智能加速器的收入份额预计将达到总收入的50%左右。 随着融合人工智能算法的服务扩展到各个行业,以及对高性价比半导体的需求不断增长,不仅传统的 无晶圆厂公司,而且云服务提供商等主要技术参与者也在积极开发针对其数据中心优化的人工智能加 ...
USB 4,更乱了
半导体行业观察· 2025-09-12 01:14
Core Viewpoint - The complexity of USB standards has become widely recognized, leading to confusion among consumers and industry professionals alike. The introduction of the USB4 standard was expected to simplify this situation, but progress has not been smooth [1]. Group 1: USB Standard Evolution - USB standards have undergone multiple updates, with USB 2.0 supporting speeds up to 480 Mbps, USB 3.x series introduced in 2008, and USB 4 series launched in 2019 with speeds up to 80 Gbps [2]. - The introduction of "Gen" symbols (e.g., Gen 1, Gen 2) has complicated the identification of USB versions, making it difficult for users to understand the maximum transfer speeds associated with each version [3][4]. Group 2: Naming Confusion - Different names can refer to the same technology, such as USB 3.0 being rebranded as USB 3.1 Gen 1, which adds to the confusion surrounding USB standards [3]. - Manufacturers have the discretion to choose naming conventions for their products, leading to inconsistencies in how specifications are presented across different brands [4]. Group 3: Market Naming Rules - The USB Implementers Forum (USB-IF) introduced market naming rules in October 2022, suggesting that names should reflect the maximum speed, such as "USB 5Gbps" for USB 3.2 Gen 1 [6]. - Despite the introduction of market names, there remains ambiguity in distinguishing between USB 4 and USB 3.x based solely on these names [7]. Group 4: Connector Standardization - USB connectors have evolved from standard A and B types to USB Type-C, which is now the standard. However, the physical shape of connectors does not indicate the speed of the USB standard [15]. - USB Type-C has become the de facto standard, supporting fast charging and display connections, but it also introduces complexity in determining data and charging directions [16]. Group 5: Ongoing Issues - The upcoming USB4 version 2.0, expected to reach speeds of 80 Gbps, will further complicate the naming conventions as it will be differentiated from previous versions [14]. - The continued use of "Gen" identifiers without specifying transfer speeds contradicts the USB-IF's goal of reducing consumer confusion [12].
Arm发布全新Lumex CSS,破局端侧AI
半导体行业观察· 2025-09-12 01:14
Core Viewpoint - The article discusses the transition of AI technology from centralized cloud computing to distributed edge deployment, emphasizing the importance of mobile devices in delivering intelligent user experiences. The launch of the Arm Lumex CSS platform is highlighted as a solution to performance bottlenecks in edge AI, enabling smarter, more efficient, and personalized experiences in consumer electronics [1][2][5]. Group 1: Industry Trends - AI technology is shifting from centralized cloud computing to distributed edge deployment, with mobile devices becoming the core carriers of intelligent experiences [1]. - The demand for low-latency, high-smoothness, and long-endurance edge AI is increasing, making edge AI a defining factor in product competitiveness [1]. - The edge computing industry faces challenges such as traditional architectures struggling to handle high-density AI tasks and increased chip design complexity leading to longer development cycles [1]. Group 2: Arm Lumex CSS Platform - Arm introduced the Lumex CSS platform, which integrates high-performance CPU, GPU, and system IP to address performance bottlenecks and development challenges in edge AI [2][5]. - The platform is designed for flagship smartphones and next-generation personal computers, aiming to optimize edge AI performance through technological innovation [7]. Group 3: Technical Innovations - The Arm C1 CPU cluster, a core component of the Lumex CSS platform, features the second-generation Scalable Matrix Extension (SME2) technology, enhancing AI workload performance by up to 5 times and energy efficiency by up to 3 times [8][10]. - The Mali G1-Ultra GPU, another key component, offers significant improvements in graphics and AI performance, including a 40% increase in game frame rates and a 20% boost in AI inference speed [18][22]. Group 4: Software Ecosystem - The KleidiAI software library is integrated with major AI frameworks, allowing developers to activate SME2 acceleration without code modifications, thus reducing development costs and barriers [26][29]. - The platform's design enables seamless integration of hardware capabilities with software, facilitating the large-scale deployment of edge AI [32][43]. Group 5: Market Impact - The global edge AI market is projected to grow from 321.9 billion yuan in 2025 to 1,223 billion yuan in 2029, with a compound annual growth rate of 39.6% [44]. - Arm's Lumex CSS platform represents a significant shift from traditional IP supplier to a full-stack solution provider, addressing industry pain points and enhancing the overall value chain [44][45].
12英寸的方形SiC晶圆曝光
半导体行业观察· 2025-09-11 01:47
目前价格部分,6吋价格下滑最严重、8吋其次,目前价格在非中的供应链稍微好一点,但也不可能都 不降价,因为如果成本太高,连客户都会失去竞争力。 公众号记得加星标⭐️,第一时间看推送不会错过。 来源 : 内容来自 经济日报 。 半导体硅晶圆厂环球晶董事长徐秀兰于开幕典礼后受访时透露,该公司将在展览中宣布开发出12吋的 方形碳化硅(SiC)晶圆,昭告该公司已具备相关能力。 徐秀兰表示,若改采用方形片,不只需要制程能力,还需要设备能力,因为并没有现成设备可用,相 关周边也要跟着变化,否则以晶圆盒来说,方形晶圆会比盒子还大,「连盒子都没得装」。或是如硅 晶圆不透光,碳化硅晶圆却透光,因此量测方面的问题也要处理。 另外,徐秀兰提到,目前市场上8吋碳化硅晶圆是以激光切割,但12吋碳化硅晶圆切割方式仍待开 发。相较于其他做碳化硅的业者,没有12吋晶圆的经验,环球晶已拥有12吋晶圆的很多经验,只是过 去是生产硅产品,而不是碳化硅,如今该公司已经开发出不用激光的12吋碳化硅晶圆切割方法。 在今年五月,碳化硅SiC晶圆大厂Wolfspeed传出声请破产,环球晶(6488)董事长徐秀兰表示,SiC是 对的材料,很多应用领域上SiC比 ...