公众号记得加星标⭐️，第一时间看推送不会错过。 如果感觉如今科技领域的一切都与人工智能息息相关，那是因为事实的确如此。而计算机内存市场更 是如此。用于为人工智能数据中心中的GPU和其他加速器供电的DRAM内存需求巨大，利润丰厚，以 至于挤占了其他用途的内存供应，导致价格飙升。据Counterpoint Research的数据显示，本季度 DRAM价格已上涨80%至90%。 最大的人工智能硬件公司表示，他们已经确保了芯片供应到 2028 年，但这让其他所有人——个人电 脑制造商、消费电子产品制造商以及其他所有需要临时存储十亿比特数据的设备制造商——都不得不 争先恐后地应对供应短缺和价格上涨的问题。 电子行业是如何陷入如今这般困境的？更重要的是，它又该如何摆脱困境？IEEE Spectrum采访了经 济学家和存储器专家，请他们对此进行解释。他们认为，如今的局面是 DRAM 行业历史上长期存在 的繁荣与衰退周期，以及规模空前的 AI 硬件基础设施建设相互碰撞的结果。而且，除非 AI 领域出 现重大崩盘，否则新增产能和新技术需要数年时间才能使供应与需求相匹配。即便如此，价格也可能 依然居高不下。 要了解事情的来龙去脉 ...

Core Viewpoint - The article discusses the advancements in Cisco's networking technology, particularly the introduction of the G300 ASIC chip, which aims to enhance data center interconnectivity and performance for AI and high-performance computing (HPC) applications [2][3][5]. Group 1: G300 ASIC Chip Features - The G300 ASIC chip offers an aggregate bandwidth of 102.4 Tb/s and is designed to compete with Broadcom and NVIDIA for market share in high-speed port connections [2][5]. - It features a unified buffer of 252 MB shared among 512 SerDes circuits, improving efficiency and reducing packet loss during network congestion [10][12]. - The chip supports various configurations, allowing for the creation of switches with different port counts and speeds, such as 512 ports at 200 Gb/s or 64 ports at 1.6 Tb/s [12][13]. Group 2: Comparison with Previous Models - The G300 is positioned as a significant upgrade over the G200, with a 33% higher network utilization and a 28% faster job completion rate [14]. - Despite potentially being three to four times the price of the G200, the G300 is considered cost-effective due to its enhanced capabilities and efficiency [21]. Group 3: Technological Innovations - The G300 utilizes a "no-cover" chip design for better heat dissipation and is manufactured using advanced processes from TSMC, including 3 nm and 4 nm technologies [7][8]. - It can directly drive linear pluggable optical modules (LPO), which can reduce power consumption by approximately 50%, contributing to a 30% reduction in overall power usage for AI infrastructure [13][21]. Group 4: Market Implications - The advancements in the G300 chip are expected to address the growing demands of AI workloads, with companies needing higher bandwidth to support upcoming GPU and XPU releases [13][21]. - Cisco's testing and certification processes for optical modules are highlighted as being more comprehensive than competitors, which is crucial for maintaining performance in AI workloads [15].

公众号记得加星标⭐️，第一时间看推送不会错过。 自 2023 年以来，数据中心的发展趋势很简单：GPU 和网络才是王道。人工智能训练和推理的出现 及其随后的爆发式增长，使得计算需求从 CPU 转向了 GPU。这意味着服务器 CPU 的主要供应商英 特尔未能搭上数据中心建设和支出的顺风车。由于超大规模数据中心和新型云平台将重心放在 GPU 和数据中心基础设施上，服务器 CPU 的收入一直相对停滞不前。 与 此 同 时 ， 这 些 超 大 规 模 数 据 中 心 运 营 商 也 纷 纷 为 其 云 计 算 服 务 开 发 基 于 ARM 架 构 的 数 据 中 心 CPU，这无疑蚕食了英特尔的一大潜在市场。而在英特尔自身的x86领域，其执行力不足、性能远逊 于竞争对手AMD，进一步蚕食了市场份额。由于缺乏有效的AI加速器产品，英特尔只能原地踏步， 眼睁睁地看着其他厂商大举扩张。 过去六个月，情况发生了巨大变化，指出CPU需求飙升。我们展示并建模的主要驱动因素是强化学习 和Vibe编码对CPU的巨大需求。 然而，英特尔近期的股价反弹以及2025年下半年不断变化的市场需求信号表明，CPU如今再次变得 至关重要。在最新 ...

公众号记得加星标⭐️，第一时间看推送不会错过。 过去几个月内存芯片价格的持续飙升，导致股市赢家和输家之间出现巨大分化，投资者看不到尽头。 从游戏机制造商任天堂、大型PC品牌到苹果公司的供应商，许多公司的股价都因盈利担忧而下跌。 与此同时，内存生产商的股价却飙升至前所未有的高度。基金经理和分析师们正在评估哪些公司能够 通过锁定供应、提高产品价格或重新设计以减少内存使用量来更好地应对供应紧张的局面。 市场此前已经有所准备：彭博社全球消费电子产品制造商指数自9月底以来下跌了12%，而包括三星 电子在内的一篮子内存制造商的股价则飙升了160%以上。问题是，市场对这种波动的消化程度如 何。 富达国际基金经理Vivian Pai表示："目前被低估的是持续时间方面的风险——目前的估值很大程度 上是基于这种波动将在1到2个季度内恢复正常的预期。"她补充说："我们认为行业供应紧张的局面可 能会持续下去"，甚至可能持续到今年年底。 内存芯片短缺和价格上涨的问题在各公司财报和电话会议中频繁出现。投资者们已经听到了警钟。 智能手机处理器制造商高通公司上周四股价下跌超过8%，此前该公司表示，内存供应紧张将限制手 机产量。任天堂在东京股 ...

人们正在寻找能够突破人工智能长期存在的"内存墙"的方法——即使是快速模型也会因处理器和内存 之间数据传输所需的时间和能量而运行缓慢。电阻式随机存取存储器（RRAM）可以通过允许计算在 内存本身进行来绕过这一障碍。然而，大多数类型的非易失性存储器都过于不稳定和笨重，无法用于 此目的。 幸运的是，可能已经有了解决方案。在12月的IEEE国际电子器件会议（IEDM）上，加州大学圣地亚 哥分校的研究人员展示了他们可以在一种全新的RRAM上运行学习算法。 "我们实际上重新设计了 RRAM，彻底重新思考了它的开关方式，"领导这项工作的加州大学圣地亚 哥分校电气工程师杜伊古·库祖姆 (Duygu Kuzum)说。 公众号记得加星标⭐️，第一时间看推送不会错过。 RRAM 将数据存储为对电流流动的电阻值。神经网络中的关键数字运算——将数字数组相乘然后将 结果相加——可以通过模拟方式轻松实现：只需让电流流过 RRAM 单元阵列，连接它们的输出，然 后测量产生的电流即可。 传统上，RRAM 通过在介电材料的高电阻环境中形成低电阻细丝来存储数据。形成这些细丝通常需 要过高的电压，超出了标准CMOS 工艺的承受范围，从而阻碍了其在处 ...

Group 1 - The core viewpoint of the article highlights TSMC's strategic shift towards advanced processes and packaging due to strong demand driven by AI, while gradually reducing its focus on mature processes [2][3] - TSMC's 8-inch annual capacity is approximately 5 million wafers, with about 80% expected to be transferred to World Advanced through various means, significantly boosting World Advanced's capacity and market share [3] - TSMC is optimizing its resource allocation by reducing some 6-inch and 8-inch wafer production while still supporting existing customer needs, indicating a strategic shift rather than a complete exit from mature processes [2] Group 2 - TSMC's investment in Arizona is transforming into a significant asset for the U.S. semiconductor industry, with plans to build up to six fabs by 2030, driven by AI demand [5][6] - The first fab in Arizona has begun large-scale production, with the second expected to be operational by 2027/2028, and a total investment exceeding $65 billion [6] - The expansion is expected to create thousands of high-tech jobs and establish Arizona as a central hub for semiconductor manufacturing in the U.S. [6][8]

站在2026年初的节点回看，半导体行业不再是一条单线叙事，由手机、PC 等单一终端主 导的传统周期，正转入以 "AI 算力基建" 为代表的多元驱动演进。 2025年，随着生成式 AI 迈入应用爆发期，全球数据中心对逻辑芯片与 HBM（高带宽存 储）的算力需求近乎狂热。以AI训练和数据中心建设为核心的新一轮算力需求，正在拉动 先进制程相关投资回暖，与此同时由AI驱动的应用也带动了对成熟制程的需求。在这一背 景下，芯片制造设备，尤其是 光刻机，可以看作衡量本轮产业复苏质量与可持续性的关键 窗口之一。 最近光刻机巨头ASML发布了2025年全年的财报。在经历了2024年的"调整年"之后，ASML 在 2025 年交出了一份极具分量的成绩单：营收、利润、在手订单全部刷新纪录。 AI 不再只是英伟达和SK海力士等芯片公司的狂欢，而是沿着台积电、英特尔、三星的产线，逐级 向上游设备传导，也体现在了ASML的订单簿上。 EUV收入增长，而DUV仍是主力 2025年，ASML实现全年净销售额约327亿欧元，全年毛利率约52.8%，净利润约96亿欧元。尽管 2023–2024年半导体行业经历资本开支波动，但最近几年ASML的营 ...

Core Viewpoint - Silicon photonics technology is transforming data centers, with significant changes expected in the future, particularly in the transition from copper cables to fiber optics for both scale-out and scale-up networking [2][4]. Group 1: Market Growth and Projections - The fiber optic device market has grown from several billion dollars in 2003 to approximately $13 billion in 2023, with projections to reach $25 billion by 2030, driven primarily by the development of AI networks [4]. - Coherent's investor report predicts that the market for pluggable optical devices will grow from $6 billion in 2023 to $25 billion by 2030, with data rates primarily at 1.6T and 3.2T [16]. - The silicon photonics wafer foundry revenue is expected to grow eightfold from 2026 to 2032, with horizontal scaling currently being the main driver [35]. Group 2: Technological Advancements - Silicon photonics integrates disparate photonic devices into improved CMOS processes, enabling higher bandwidth and lower power consumption compared to traditional copper cables [10][16]. - The transition from copper to fiber optics is facilitated by pluggable optical transceivers, which connect to electrical interfaces on switches or servers, allowing for high-speed data transmission [16]. - Co-packaged optics (CPO) are emerging as a more efficient alternative to pluggable optical devices, offering higher density and lower power consumption [21]. Group 3: Key Players and Innovations - Major players in the silicon photonics foundry market include GlobalFoundries, Tower Semiconductor, and TSMC, with TSMC expected to become the leading foundry due to its extensive capabilities in AI accelerator production [28][37]. - Companies like Nvidia and Broadcom are set to launch Ethernet switches using co-packaged optical devices by 2025, indicating a shift in market dynamics [21]. - Startups such as iPronics, nEye, and Salience are developing compact silicon photonics technologies for optical circuit switching systems, which may offer more economical and reliable solutions [20]. Group 4: Challenges and Future Directions - Signal loss remains a significant challenge in silicon photonics, necessitating precise control over signal integrity during design [45]. - The integration of silicon photonics with CMOS technology is still in its early stages, but advancements are expected to bring more structure and foundational knowledge to the field [39]. - The industry is likely to see a transformation in manufacturing structures, with TSMC poised to leverage its experience in AI accelerator manufacturing to become a dominant player in silicon photonics [53].

Core Insights - The rapid development of artificial intelligence (AI) is driving increased attention towards foundries producing Power Management Integrated Circuits (PMIC) as market demand diversifies into areas like data centers and electric vehicles (EV) [2] - DB HiTek, a South Korean 8-inch wafer foundry, projects revenues and operating profits of 1.4 trillion KRW (approximately 954.85 million USD) and 277.3 billion KRW respectively for 2025, reflecting year-on-year growth of 24% and 45% [2] - Despite a decline in revenue for 2023 and 2024 due to weak IT equipment demand, DB HiTek anticipates a rebound in 2025 driven by a recovery in power semiconductor demand and growth in AI and EV markets [2] - The average utilization rate of DB HiTek's wafer factories is expected to rise significantly from 76% in 2024 to 96% in 2025 [2] Industry Trends - The demand for 8-inch wafers is continuing to grow despite major foundries like Samsung Electronics and TSMC reducing their 8-inch wafer production capacity [3] - The automotive voltage systems are transitioning from traditional 12 volts to 48 volts, while AI data centers are increasing operational voltages from 380 volts to as high as 800 volts, necessitating technologies capable of handling higher voltages [3] - DB HiTek plans to expand its business through high-voltage process technology, while SK Hynix's subsidiary SK Keyfoundry aims to strengthen its position in the PMIC market by introducing new high-voltage processes and collaborating closely with customers on product development [3]

公众号记得加星标⭐️，第一时间看推送不会错过。 四年前，英特尔开始为其第四代至强处理器筹备软件定义芯片（SDSi）计划，旨在销售处理器的同 时 ， 通 过 允 许 客 户 激 活 特 定 工 作 负 载 的 加 速 器 来 赚 取 溢 价 。 最 终 ， 该 计 划 正 式 更 名 为 "Intel On Demand"（英特尔按需服务） ……然后便销声匿迹。如今，据Phoronix报道，这家标志性的 PC 芯 片制造商正在悄然终止该计划。 近年来，英特尔基本停止了公开讨论其"Intel On Demand"计划，而缺乏新的补丁也表明其开发活动 放缓，这可能意味着英特尔已不再将其视为优先事项。Phoronix最新报道称，包含支持 Intel On Demand 所需软件组件的 Intel SDSi GitHub 代码库已于去年 11 月被存档，这标志着该计划的积极 开发已经结束。与此同时，英特尔也从其网站上删除了大部分 On Demand 相关文档，目前所有与该 计划相关的内容都只能通过其网站访问到一些旧的 PDF文档。 综上所述，软件支持、文档和公开讨论的消失强烈表明，英特尔已经彻底放弃了软件定义芯片计划， ...