Workflow
半导体行业观察
icon
Search documents
台积电买了30台EUV光刻机
半导体行业观察· 2025-10-23 01:01
来 源: 内容来自工商时报 较新的 2 纳米制程可以继续利用台积电现有的 EUV 光刻机进行量产,并提高良率。但随着这家台湾 半导体巨头向 1.4 纳米和 1 纳米(也称为 A14 和 A10)等 2 纳米以下节点迈进,它将面临制造方面 的障碍。现在,通过购买 ASML 先进的高数值孔径 EUV 光刻机可以轻松解决这个问题,但一份新 报告指出,台积电将不再购买 ASML 光刻机,而是转向光掩模薄膜。 2纳米晶圆预计将于2025年底全面投产,之后台积电将最终过渡到1.4纳米节点。该公司已为 迈向先 进制程奠定了基础,预计将于2028年左右开始量产。台积电初期投资高达1.5万亿新台币(约合490 亿美元),目前正以闪电般的速度推进,在新竹工厂启动1.4纳米制程的研发,并购置了30台EUV光 刻机。 然而,台积电拒绝采取的一项举措是购买ASML高数值孔径EUV光刻机,单台售价4亿美元。然而, 这台设备将确保以更高的良率可靠地生产1.4纳米和1纳米晶圆。该公司可能认为,这台设备的价值与 其实际价值不符,因此,据Dan Nystedt和《商业时报》报道,台积电正在转向光掩模防护膜。亚2纳 米工艺必须使用防护膜,以防止灰 ...
玻璃基板,大势所趋
半导体行业观察· 2025-10-23 01:01
公众号记得加星标⭐️,第一时间看推送不会错过。 先进封装中的玻璃如今已成为平台业务,而非零部件业务。对于玻璃载体而言,收入来源已从板价转 向单程价格,其经济效益取决于重复使用次数、激光/UV 脱键产量、良率以及边缘损伤避免。这为提 供 CTE 等级系列的供应商、将载体 + 粘合剂/LTHC + 脱键作为单一合格堆栈进行销售的捆绑包拥有 者以及拥有光学质量保证的区域回收供应商带来了短期收益。拥有深厚玻璃专业知识的公司(例如, Plan Optik 的高平整度、可控传输且边缘经过精心设计的载体捆绑销售的模式)处于最佳位置。玻璃 芯基板通过增加TGV(超大孔径风洞)+精细RDL(重布线层)+积层工艺,将显示面板产能转化为 利润。赢家占据着关键的接口:高良率TGV钻孔/蚀刻技术,无空洞铜填充,自适应套刻的面板光刻 技术,2/2 μm L/S(线宽/线厚)以及可控制翘曲的面板处理技术。与显示玻璃制造商合作的基板厂 商和OSAT厂商将面积产能转化为大尺寸封装的成本优势。 玻璃已从简单的载体发展成为先进封装的完整材料平台,顺应了芯片集、面板化、垂直集成和混合键 合等大趋势,同时缩紧了机械、热和洁净度方面的预算。作为载体(晶 ...
韩国芯片出口,创新高
半导体行业观察· 2025-10-23 01:01
公众号记得加星标⭐️,第一时间看推送不会错过。 来 源: 内容来自半导体行业观察综合。 南韩产业通商资源部第一副部长文慎鹤周三(22 日) 表示,由于全球人工智能(AI) 市场不断扩大,对 先进半导体的需求日益增长,南韩芯片出口额预计将在2025 年连续第二年创下历史新高。 文慎鹤在第18 届半导体日活动上指出,今年南韩半导体出口额有望超越1650 亿美元。这项预期是建 立在去年半导体出口额达1419 亿美元的基础上。 根据政府数据,今年1 月至9 月期间,南韩出口的半导体价值已达1197 亿美元,较去年同期增长了 16.9%。 面对芯片产业的强劲表现,南韩政府承诺将持续努力,以协助南韩半导体产业在记忆体芯片市场中保 持主导地位。同时,政府也计画积极缩小与全球领先企业在其他领域,例如系统半导体和无晶圆厂 (fabless) 技术方面的差距。 南韩半导体产业协会会长Song Jai-hyuk 对此强调,半导体是「国家战略资产」,在AI 和量子计算 领域发挥着至关重要的作用。他呼吁政府对此产业提供「积极支持」,并致力于建立一个「创新的」 产业生态系统。 尽管出口前景看好,但在周三上午的南韩股市交易中,两大芯片巨头 ...
量子芯片,史上首次,谷歌造
半导体行业观察· 2025-10-23 01:01
Core Insights - Google has achieved a historic milestone by successfully running a verifiable quantum algorithm that surpasses the capabilities of the fastest classical supercomputers, demonstrating a speed increase of 13,000 times [2][5] - The breakthrough is based on the Willow quantum chip, which significantly reduces error rates and enhances computational speed, paving the way for practical applications in fields like medicine and materials science [6][11] Quantum Echoes Algorithm - The Quantum Echoes algorithm represents the first verifiable quantum advantage, meaning results can be replicated on Google's quantum computer or any equivalent quantum system, confirming the outcomes [5][6] - This algorithm simulates physical experiments and tests both complexity and accuracy, making it a significant step towards practical quantum computing applications [6] Willow Quantum Chip - The Willow chip has achieved two major advancements: it reduces errors exponentially with an increase in qubit numbers and completes a standard benchmark calculation in five minutes, a task that would take the fastest classical supercomputer approximately 10^25 years [11][15] - The chip features 105 qubits and has shown impressive performance in quantum error correction and random circuit sampling benchmarks, indicating its potential for scalable quantum computing [18][19] Applications in Science - Quantum computing is expected to play a crucial role in modeling quantum mechanical phenomena, such as atomic interactions and molecular structures, which are foundational in chemistry, biology, and materials science [8][9] - Experiments conducted in collaboration with UC Berkeley using the Quantum Echoes algorithm on the Willow chip have validated the method against traditional nuclear magnetic resonance (NMR) results, revealing new insights [8] Future Prospects - Google aims to demonstrate practical, super-classical computations that are relevant to real-world applications, bridging the gap between quantum computing capabilities and commercial viability [19] - The advancements in quantum error correction and the performance of the Willow chip suggest that practical large-scale quantum computers are becoming increasingly feasible [13][17]
大众预警:Nexperia 芯片供应中断,生产或中断
半导体行业观察· 2025-10-23 01:01
Core Viewpoint - The recent takeover of Nexperia by the Dutch government has led to export restrictions from China, causing potential production disruptions in the automotive industry, particularly affecting companies like Volkswagen and General Motors [2][4][5]. Group 1: Impact on Automotive Industry - Volkswagen has warned of possible temporary production halts due to export restrictions on semiconductors produced by Nexperia, despite not being a direct supplier [2][4]. - The German Automotive Industry Association (VDA) has indicated that if the chip supply disruption is not resolved quickly, it could lead to severe production limitations [2][6]. - General Motors has formed an internal team to mitigate potential disruptions from the Nexperia situation, emphasizing the current instability of the situation [4][5]. Group 2: Nexperia's Situation - The Dutch government intervened in Nexperia's operations citing serious governance issues and concerns over economic security risks [5][6]. - Nexperia has notified its clients that it cannot guarantee the supply of chips to the automotive supply chain, raising alarms among manufacturers [6][7]. - The company has been under scrutiny due to its ties with the Chinese firm Wingtech Technology, which has faced export restrictions from the U.S. [7][8]. Group 3: International Reactions - The Chinese government has reacted by imposing export bans on certain products from Nexperia, leading to heightened tensions between China and the Netherlands [3][5]. - Discussions between Chinese and Dutch officials are ongoing, aiming to find a constructive solution to the semiconductor supply chain issues [8][10]. - The situation has escalated into a broader technology dispute between China and the West, impacting global supply chains [4][10].
一家芯片初创公司,单挑Nvidia和Intel
半导体行业观察· 2025-10-23 01:01
公众号记得加星标⭐️,第一时间看推送不会错过。 在2024 年 10 月退出隐身模式的时候,以色列芯片初创公司NextSilicon 表示,期即将退出的 Maverick-2 是世界上第一款智能计算加速器 (ICA:Intelligent Compute Accelerator),旨在满 足高性能计算人工智能 (HPC-AI) 应用的需求,是一种"新颖且原创的计算架构",可在降低功耗 和成本的同时提高性能。 刚刚。经过八年时间、3.03 亿美元的种子资金和三轮风险投资的NextSilicon 终于推出了其 64 位数 据流引擎的多个版本。与此同时,该公司还将推出一款名为 Arbel 的自主研发 RISC-V 处理器,该 芯片或将与 Maverick-2 搭配使用,打造诸如英伟达"Superchip"类型的产品。 从左到右: NextSilicon Arbel RISC-V CPU 、 Maverick-1 DFP 、 Maverick-2 DFP 和用于 OAM 插座的双 芯片 Maverick-2 。 NextSilicon 成立于 2017 年,远早于 GenAI 热潮兴起之时,但当时人们已经意识到 HP ...
三星HBM4,首次亮相
半导体行业观察· 2025-10-23 01:01
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 wcctech 。 三星首次向公众展示了其 HBM4 内存模块,这表明这家韩国巨头确实为即将到来的 HBM 竞争做好 了准备。 虽然目前市场以第五代 HBM3E 芯片为主,但业内观察人士预计 HBM4 将成为明年的主要因素,因 为 Nvidia 计划在其下一代 AI 加速器 Rubin 中使用它。 SK海力士目前是HBM3E的主要供应商,与Nvidia和台积电组成了三方供应链,目前已完成HBM4的 开发,并正在准备量产。据报道,该公司正在与Nvidia洽谈大规模供应事宜。 据报道,三星正在避免重蹈覆辙,避免在 DRAM 领域失去主导地位。为了确保不落后,这家韩国巨 头正与竞争对手一起推进 HBM4 的量产。据《电子时报》报道,三星 HBM4 逻辑芯片的良率已达到 惊人的 90%,这表明该公司的量产进度已步入正轨,更重要的是,目前预计不会出现延期。 据报道,这家韩国巨头还在实施多项策略,以确保HBM4的早期普及,包括保持有竞争力的价格、提 供更高的产能,更重要的是,为NVIDIA等客户提供更快的引脚速度(额定速度约为11 Gbps),高 于 ...
十篇论文,揭秘寒武纪AI芯片崛起之路
半导体行业观察· 2025-10-23 01:01
Core Insights - The article discusses the rise of Cambricon, a leading AI chip company in China, highlighting its technological evolution and competitive edge against global giants like NVIDIA [5][26]. Group 1: Foundational Era - The inception of Cambricon is attributed to the academic journey of two brothers, Chen Yunji and Chen Tianshi, who laid the groundwork for deep learning processor architecture through their research at the Chinese Academy of Sciences [7]. - The "DianNao" series, introduced by the brothers, was one of the earliest systematic studies on deep learning processor architectures, addressing the efficiency bottlenecks of general-purpose CPUs/GPUs in executing neural networks [7][12]. Group 2: Technological Evolution - The article highlights ten significant papers published between 2014 and 2025, tracing the technological advancements from the "DianNao" architecture to the Cambricon series of AI chips [5]. - The first paper, "DianNao," demonstrated a high-throughput accelerator capable of executing 452 GOP/s with a power consumption of 485 milliwatts, achieving a speedup of 117.87 times compared to a 128-bit 2GHz SIMD processor [11]. - Subsequent innovations, such as "DaDianNao" and "PuDianNao," showcased significant performance improvements, with "DaDianNao" achieving a 450.65 times speedup over GPUs and "PuDianNao" supporting seven mainstream machine learning algorithms [14][20]. Group 3: Commercialization and Ecosystem Development - Cambricon's transition from academic research to commercial products was marked by the introduction of the "Cambricon ISA," a specialized instruction set for deep learning, which decoupled upper applications from lower hardware [26][30]. - The integration of Cambricon-1A into Huawei's Kirin 970 chip marked a significant commercial breakthrough, establishing Cambricon as a key player in the mobile AI chip market [37]. - Following the loss of Huawei as a major client, Cambricon pivoted to focus on its "Siyuan" (MLU) cloud chips and the NeuWare software platform, aiming to compete with NVIDIA's ecosystem [37]. Group 4: Future Challenges and Opportunities - The article concludes by emphasizing the challenges Cambricon faces against NVIDIA's established technology and the need to carve out a unique path in the AI chip market [59]. - Despite the challenges, the growing demand for autonomous AI computing in China presents a significant opportunity for Cambricon to leverage its academic roots and build a robust developer ecosystem [59].
干掉40%的工程师?初创公司推动AI开发芯片
半导体行业观察· 2025-10-22 01:20
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forbes 。 从最新的消费电子产品到人工智能 (AI) 的新应用,在推出先进技术的竞争中,一项挑战常常会阻碍 竞争对手。将这些创新核心的微芯片理论转化为可直接安装的制造组件,需要长达四年的时间。随着 芯片复杂性的不断提高,这一挑战只会变得更加艰巨。 " 我 们 迫 切 需 要 看 到 芯 片 开 发 周 期 的 加 速 , " 瑞 士 初 创 公 司 Chipmind 的 联 合 创 始 人 兼 首 席 执 行 官 Harald Kröll 表示。"技术进步正受到其所依赖的芯片开发过程的阻碍,该过程复杂、耗时且成本高 昂。" 本轮融资由瑞士种子基金 Founderful 领投。Founderful 负责人 Edouard Treccani 表示:"在这个人 工智能日渐普及的世界里,Chipmind 脱颖而出,为 Harald 和 Sandro 潜心钻研 20 年的难题提供了 令人耳目一新的解决方案。" 利用代理 AI 改造半导体设计 Chipmind 今天宣布完成 250 万美元的种子轮前融资,该公司相信自己有能力提供帮助。这家总部位 ...
本土激光雷达大厂CEO:特斯拉纯视觉方案不够安全
半导体行业观察· 2025-10-22 01:20
Core Viewpoint - The founder of Chinese LiDAR manufacturer RoboSense, Qiu Chunchao, argues that multi-sensor systems are superior and safer for autonomous vehicles compared to the pure vision system promoted by Tesla's CEO Elon Musk [2][4]. Group 1: LiDAR vs. Vision Systems - LiDAR, which stands for Light Detection and Ranging, is a sensor technology that scans the environment by emitting laser beams and measuring the time it takes for the signals to return [2]. - Qiu emphasizes that relying solely on vision systems is insufficient for achieving Level 3 or Level 4 autonomous driving capabilities, necessitating the inclusion of additional sensors like LiDAR [2][3]. - Market research firm Yole Group predicts that RoboSense will capture the largest market share of global passenger car LiDAR systems by 2024 [3]. Group 2: Musk's Perspective on LiDAR - Musk has been a long-time critic of LiDAR systems, asserting that the future of autonomous driving lies solely in the use of cameras [4][6]. - He claims that the reliance on cameras is the most "human-like" way to navigate, as humans use their eyes for navigation [6]. - The cost of LiDAR systems is significantly higher, approximately $12,000 per vehicle, compared to around $400 for cameras [4][6]. Group 3: Industry Opinions on Sensor Technology - Other companies like Waymo and Zoox utilize a combination of cameras and sensors, including radar and LiDAR, to enhance object detection in adverse weather and low-light conditions [5]. - Uber's CEO Dara Khosrowshahi supports the use of a combination of sensors, including LiDAR, for achieving superior safety in autonomous vehicles [6][7]. - Qiu points out that the cost of LiDAR systems has dramatically decreased from around $70,000 per vehicle to a few hundred dollars, while performance has improved [7]. Group 4: Regional Differences in Autonomous Driving - Li Xiang, CEO of Chinese electric vehicle manufacturer Li Auto, suggests that Musk's dismissal of LiDAR may stem from differences in traffic conditions between the U.S. and China [7][8]. - He argues that in China, drivers often encounter poorly lit or malfunctioning vehicles, which current camera systems may struggle to detect [8].