来源：内容来自工商时报，谢谢。 如果您希望可以时常见面，欢迎标星收藏哦~ 美国总统川普最快本周宣布对进口芯片实施关税，市场预估税率可能介于25％～100％；法人指 出，此芯片关税恐对台积电及台系半导体业者产生深远影响，新规则不排除以晶圆制造地（wafer out）就源课税，半导体业者皆屏息以待。 业者表示，台湾半导体产业被视为「硅盾」，但川普芯片关税政策可能削弱这一优势。若关税过 高，美国客户可能转向其他国家（例如南韩或日本）寻求替代供应商，或促使中国加速自制芯片进 程。同时，美国对台积电的依赖与地缘政治紧张局势，若加速执行「美国制造」目标，将使关税政 策可能引发报复性措施，进一步扰乱全球供应链。 台系IC设计业者近日于法说会上强调，4月初执行的对等关税并未见重大订单调整，客户下单多采 取稳定态度观望面对，惟对于半导体关税，业者认为，采取晶圆产地课税，将使得美国客户面临进 口关键零组件成本显著提高，首先冲击的就是于亚洲投片生产的美国芯片业者。 若新制如市场推测，改以wafer out认定课税，不排除推动部分产能回流美国本土，但预计供应链 重组过程中，将出现采购周期拉长、运输递延等不确定性，必定影响企业获利，同 ...

Core Viewpoint - TSMC's 2nm process technology is advancing rapidly, showing defect density performance comparable to the 5nm family and surpassing the 7nm and 3nm processes, indicating a high level of technological maturity and strong potential for mass production in the second half of 2025 [1][2]. Group 1: Technological Advancements - TSMC's 2nm process is expected to significantly boost revenue for supply chain partners such as Zhonghua and Shengyang [1]. - The 2nm process will utilize GAAFET architecture, replacing the long-standing FinFET technology, which enhances transistor density and efficiency while reducing leakage current and power consumption [1]. Group 2: Clientele and Market Demand - Major clients for TSMC's 2nm process include leading companies like Apple, NVIDIA, AMD, Qualcomm, MediaTek, and Broadcom [1]. - TSMC's chairman emphasized that the demand for 2nm technology is unprecedented, far exceeding that of the 3nm process [1]. Group 3: Production Capacity and Expansion Plans - TSMC's 2nm process has entered trial production, with plans to start engineering line validation at the Hsinchu Baoshan Fab 20 in Q4 2024, targeting a monthly capacity of 3,000 wafers, and aiming for mass production by Q4 2025 with a capacity of 30,000 wafers per month [2]. - By 2027, TSMC plans to expand the total monthly capacity for 2nm at Hsinchu and Kaohsiung to 120,000 to 130,000 wafers, with a potential to reach 50,000 wafers by the end of 2025 [2]. - TSMC is investing over 1.5 trillion NTD to accelerate the construction of additional fabs in Hsinchu and Kaohsiung, establishing the largest semiconductor manufacturing hub globally [2].

如果您希望可以时常见面，欢迎标星收藏哦~ 来源：内容编译自miragenews ，谢谢 。 英国伦敦玛丽女王大学、诺丁汉大学和格拉斯哥大学的科学家团队获得了一项价值600万英镑的 EPSRC项目资助，该项目名为"利用超低能耗二维材料和器件（NEED2D）实现净零排放和人工智 能革命"。该项目将开发节能、原子级厚度的半导体，以大幅降低人工智能数据中心和高性能计算 的电力需求。 该团队由伦敦玛丽女王大学牵头，将与众多制商和多家研究机构（超过20个合作伙伴为该项目贡 献超过200万英镑）合作，开发新材料并制造晶体管等革命性的低能耗电子设备原型。这将使英国 能够利用创新的二维半导体，打造一个超越传统材料的全新电子产业。 该项目负责人、伦敦玛丽女王大学材料科学教授科林·汉弗莱斯爵士表示："世界各国政府正斥资数 十亿美元建设风能、太阳能、核能和天然气发电站，以满足人工智能数据中心巨大的能源需求。我 们的方法是从源头上解决问题：首先减少这些中心的能源消耗。" 英国已经拥有欧洲最大的数据中心市场，伦敦是其中心枢纽。用低功耗二维晶体管取代高能耗的硅 芯片，将有助于确保英国继续成为吸引科技投资的地方，同时证明能源转型的经济潜力。 ...

Core Viewpoint - The article highlights the challenges in semiconductor parameter extraction due to the complexity of device models and the inefficiencies of traditional optimization algorithms. It introduces Keysight's ML Optimizer, a machine learning-based global optimizer that significantly improves the parameter extraction process, reducing the time from days to hours and enhancing accuracy and consistency in model fitting [1]. Group 1: Challenges in Semiconductor Parameter Extraction - The complexity of semiconductor device models has made parameter extraction increasingly challenging [1]. - Traditional optimization algorithms struggle with unclear gradient changes, often getting trapped in local optima, leading to unsatisfactory extraction results [1]. - The presence of numerous interrelated parameters in modern semiconductor models further complicates the efficiency of traditional methods, requiring engineers to break down the extraction process into lengthy sub-steps [1]. Group 2: Introduction of ML Optimizer - Keysight has launched the ML Optimizer, which utilizes machine learning to provide a revolutionary solution for semiconductor parameter extraction [1]. - The ML Optimizer can process vast amounts of data and parameters in a single step, greatly simplifying the extraction workflow [1]. - The time required for parameter extraction is reduced from several days to just a few hours, significantly enhancing work efficiency [1]. Group 3: Advantages of ML Optimizer - The ML Optimizer excels in navigating non-convex parameter spaces, overcoming the limitations of traditional methods [1]. - It employs advanced machine learning algorithms to more accurately identify global optima, improving the precision of parameter extraction [1]. - The overall consistency of model fitting is enhanced, providing a solid foundation for the accurate construction of semiconductor device models [1].

Core Viewpoint - The rapid development of artificial intelligence (AI) technology has significantly boosted the demand for high bandwidth memory (HBM), contributing to SK Hynix's record performance last year and highlighting its role in leading technological changes in the AI era [1][4]. Group 1: Strategic Vision and Leadership - The driving force behind the HBM business planning organization is a sense of "pride," which will inject new vitality into its development and lead the organization towards greater growth goals [2]. - The newly appointed leader, Vice President Choi Jun-Long, emphasizes the importance of teamwork and aims to transform SK Hynix into a "Full Stack AI Memory Provider" by delivering customized HBM products that meet diverse customer needs [2][4]. - Choi Jun-Long has successfully led the delivery of the sixth-generation HBM product, "12-layer HBM4," establishing a competitive advantage in the global HBM market [3]. Group 2: Market Demand and Production Challenges - The semiconductor demand has reached unprecedented heights due to the AI boom, with HBM being the most suitable product for power efficiency and performance requirements [4]. - SK Hynix is committed to maintaining its leading position in the HBM market by advancing the mass production of the 12-layer HBM4 and responding to customer needs with HBM4E [4][6]. - The company faces challenges in scaling production lines to meet the surging demand for HBM products, particularly in light of the rapid growth in AI applications [5][6]. Group 3: Innovation and Collaboration - The new Vice President of HBM Heterogeneous Integration Technology, Han Kwon-Hwan, highlights the importance of both technological and operational innovation to respond to market demands effectively [5][6]. - The focus is on building a collaborative system that can quickly respond to market and customer needs, ensuring stable mass production [6][7]. - Enhancing production line flexibility and fostering closer cooperation with customers are key strategies to maximize the competitive advantage of SK Hynix's HBM products [7].

Core Viewpoint - In April, South Korea's exports reached a historical high of $58.21 billion, marking a year-on-year increase of 3.7%, while imports decreased by 2.7% to $53.32 billion, resulting in a trade surplus of $4.88 billion [1][2]. Group 1: Export Performance - South Korea's exports have shown positive growth for three consecutive months, driven primarily by semiconductor shipments, which increased by 17.2% year-on-year to $11.7 billion, setting a record for April [1][2]. - The export of automobiles decreased by 3.8% year-on-year, yet still reached $6.5 billion, the highest monthly export figure for the year so far, indicating resilient overall demand [1][2]. - Agricultural and marine product exports reached $1.1 billion, achieving the highest monthly export record for this category, influenced by the global popularity of Korean food (K-food) [1]. Group 2: Import Trends - Total imports in April amounted to $53.32 billion, reflecting a year-on-year decline of 2.7%, with energy imports dropping by 20.1% to $10 billion [2]. - Non-energy imports, including semiconductor manufacturing equipment, grew by 2.4% to $43.4 billion, with semiconductor manufacturing equipment alone increasing by 18.2% [2]. Group 3: Trade Surplus and Market Insights - The trade surplus for April reached $4.88 billion, an increase of $3.6 billion compared to the same month last year, with a cumulative trade surplus of $12.2 billion for the first four months of the year, up by $2.3 billion year-on-year [2]. - The Korean Trade Minister highlighted that despite a decline in exports to the U.S., strong performance in other major markets has helped maintain a positive growth trend [2].

如果您希望可以时常见面，欢迎标星收藏哦~ 来源：本文编译自entrepreneur，谢谢。 英伟达首席执行官黄仁勋可能因其持有Nvidia 3.4 % 的股份而成为亿万富翁，但他最近才获得加 薪——而过去十年他的薪水一直没有变化。 这位 62 岁的首席执行官目前的总薪酬是 Nvidia 员工平均薪酬的 166 倍。 从 2014 财年到 2024 财年，黄仁勋每年的薪水接近 100 万美元，约为 996,800 美元。根据周四 提交给美国证券交易委员会的最新代理文件，在截至 2025 年 1 月 26 日的 2025财年，黄仁勋的 基本工资上涨了 50%，达到 150 万美元。 英伟达薪酬委员会在文件中指出，这是"黄仁勋10年来首次加薪"，考虑到同类公司高管的基本工 资，此举"是合适的"。例如，AMD首席执行官苏姿丰（Lisa Su）自2024年7月起的基本工资为 126万美元。 黄在2025财年的总薪酬为4990万美元。其中包括黄作为非股权激励计划的一部分获得的600万美 元、3880万美元的股票奖励，以及350万美元的住宅安保、司机服务、安全监控和汽车费用。 文件指出，2025 财年 Nvidia 员工的 ...

如果您希望可以时常见面，欢迎标星收藏哦~ 来源：本文编译自tomshardware，谢谢。 英特尔正在测试一种新方法，以应对其高耗电芯片不断增长的发热量。在最近的 Foundry Direct Connect 活动上，该公司展示了一种实验性的封装级水冷解决方案，旨在更高效地冷却 CPU。英 特尔拥有 LGA（平面栅格阵列）和 BGA（球栅阵列）两种 CPU 的工作原型，并使用英特尔酷睿 Ultra和至强服务器处理器进行了演示。 该冷却解决方案并非将冷却液直接喷洒在硅片上，而是在封装顶部放置一个专门设计的紧凑型冷却 块，其特征是铜制微通道，可精确引导冷却液流动。这些通道可进行优化，以针对硅片上的特定热 点，从而有可能在最关键的部位改善散热效果。 对于业内人士来说，这可能一开始没有意义，但看看未来的设计，在不同的工艺节点上可能会有不 同类型的tiles并执行不同的功能，有些区域比其他区域更热是有道理的，特别是因为每个tiles都可 以与周围的tiles相互作用。 目前，英特尔代工厂拥有多种热界面材料。"新型TIM"听起来像是液态金属。某些类型的TIM更适 合不同的应用，因此存在不同的选择也是合情合理的。 | Hi ...

Core Insights - The advancement of artificial intelligence (AI) relies on the exponential growth of AI supercomputers, with training compute power increasing by 4.1 times annually since 2010, leading to breakthroughs in various AI applications [1][13] - The performance of leading AI supercomputers doubles approximately every nine months, driven by a 1.6 times annual increase in the number of chips and their performance [2][3] - By 2025, the most powerful AI supercomputer, xAI's Colossus, is estimated to have a hardware cost of $7 billion and a power demand of around 300 megawatts, equivalent to the electricity consumption of 250,000 households [3][41] Group 1: AI Supercomputer Performance and Growth - The performance of leading AI supercomputers is projected to grow at an annual rate of 2.5 times, with private sector systems growing even faster at 3.1 times [21][29] - The number of AI chips in top supercomputers is expected to increase from over 10,000 in 2019 to over 200,000 by 2024, exemplified by xAI's Colossus [2][24] - The energy efficiency of AI supercomputers is improving, with a yearly increase of 1.34 times, primarily due to the adoption of more energy-efficient chips [45][49] Group 2: Hardware Costs and Power Demand - The hardware costs of leading AI supercomputers are projected to double annually, reaching approximately $2 billion by 2030 [50][73] - Power demand for these supercomputers is expected to grow at a rate of 2.0 times per year, potentially reaching 9 gigawatts by 2030, which poses significant challenges for infrastructure [41][75] - The rapid increase in power demand may lead companies to adopt distributed training methods to manage workloads across multiple locations [76][77] Group 3: Market Dynamics and Geopolitical Implications - The private sector's share of AI supercomputer performance has surged from under 40% in 2019 to about 80% by 2025, while the public sector's share has dropped below 20% [8][56] - The United States dominates the global AI supercomputer landscape, accounting for approximately 75% of total performance, followed by China at 15% [10][59] - The shift from public to private ownership of AI supercomputers reflects the growing economic importance of AI and the increasing investment in AI infrastructure [54][68]

Core Viewpoint - Photonics is becoming increasingly essential in accelerating artificial intelligence in data centers, with over 60% of the optical components market now driven by AI data communication [2][37]. Group 1: Market Overview - The global optical components market reached $17 billion last year, historically dominated by the telecom sector, but now significantly influenced by AI-driven data centers [2][4]. - Major suppliers in the optical components market include Coherent and Aisino Technology, each holding a 20% market share, followed by Broadcom with 10% [4]. Group 2: AI and Data Center Growth - Large Language Models (LLMs) are driving exponential growth in AI workloads, necessitating vast XPU clusters and high-bandwidth, low-latency network solutions [5][6]. - Network costs in data centers are projected to rise from 5%-10% of capital expenditures to 15%-20% by 2030 [5]. Group 3: Network Architecture - Two primary types of interconnects in AI data centers are horizontal scaling (fiber links connecting switches across racks) and vertical scaling (electrical links connecting GPUs within racks) [7][10]. - The transition from copper to photonics in vertical scaling networks is ongoing but not yet complete [12]. Group 4: Power and Efficiency - The shift to higher data rates in horizontal scaling networks is leading to increased power consumption, with Nvidia reporting a reduction in power from 30W to 9W when transitioning from pluggable optical modules to Co-Packaged Optics (CPO) [18]. - CPO technology can significantly enhance GPU density, potentially increasing the number of GPUs by up to three times within the same power budget [18]. Group 5: Reliability and Challenges - Reliability is a critical factor in the transition from copper to fiber optics, with a low failure rate essential for handling the vast amounts of data generated in AI data centers [21]. - Nvidia's roadmap indicates that while current solutions use copper, the increasing data rates and signal integrity issues will necessitate a shift to fiber optics [26]. Group 6: Future Outlook - The market for CPO is expected to grow from zero to $5 billion by 2030, with early entrants like Broadcom, Marvell, and Ayar Labs poised to benefit [33]. - By the mid-2030s, all interconnects are anticipated to be optical, utilizing CPO technology [37].