Workflow
半导体芯闻
icon
Search documents
台积电A16 首发,唯一合作客户曝光
半导体芯闻· 2025-12-02 10:35
如果您希望可以时常见面,欢迎标星收藏哦~ 根据Wccftech 报导,传英伟达极可能成为台积电A16 制程(1.6 纳米)的唯一客户,并已将此 技术锁定于英伟达新一代GPU「Feynman」。 供 应 链 消 息 显 示 , 英 伟 达 的 Rubin 与 Rubin Ultra 系 列 将 率 先 采 用 3 纳 米 , 而 再 下 一 代 的 Feynman 则计画直接跨入A16。为配合此时程,台积电高雄P3 厂正加速建置,预计在2027 年 替英伟达启动量产。近期台积电扩充3 纳米产能,也被业界解读为因应英伟达大量拉货、并提 前为A16 布局。 A16 采用纳米片电晶体架构,并搭配SPR 背面供电技术,可释放更多正面布局空间、提升逻辑 密度并降低压降,其背面接面(Backside Contact) 亦能维持传统版图弹性,是业界首创的背 面供电整合方案。相较N2P,A16 在相同电压下速度提升8–10%,相同速度下降低15–20% 功 耗,芯片密度提升至1.10 倍,特别适合AI 与HPC 等高运算密度芯片。 除了台积电外,其他晶圆大厂也正加速布局背面供电技术。三星已在今年的晶圆代工论坛宣 布,将于202 ...
英特尔先进封装发威
半导体芯闻· 2025-12-02 10:18
如果您希望可以时常见面,欢迎标星收藏哦~ 据韩国媒体etnews报道,英特尔正在其位于仁川松岛的Amkor工厂(晶圆厂)推进人工智能 (AI)半导体封装技术。AI封装技术此前一直由英特尔在其自有晶圆厂独家研发,此次是英特 尔首次将该工艺外包。英特尔选择韩国晶圆厂作为强化其半导体供应链的战略基地,这一决定 值得关注。 据业内人士1日透露,英特尔已在Amkor松岛K5工厂建立了尖端封装技术"EMIB"工艺。英特尔 和Amkor于今年4月签署了EMIB技术合作协议,松岛K5工厂被选为实际合作的实施地点。 EMIB是一种2.5D封装技术,可将不同的半导体(芯片)连接起来。例如,在AI加速器中,图 形处理器(GPU)位于中心,高带宽内存(HBM)环绕其周围。处理器和内存之间的信号通过 名为"EMIB(嵌入式多芯片互连桥)"的路径传输,这也是其名称的由来。 目前,该路径采用硅中介层实现。英伟达的AI加速器就是一个典型的例子。然而,硅中介层成 本高昂。EMIB利用嵌入半导体衬底中的硅桥,据称与硅中介层相比,具有更高的性价比和生产 效率。它还拥有精确的2.5D封装优势。 英特尔晶圆代工旨在利用领先的制程工艺技术,为英特尔内部 ...
MLCC,需求大增
半导体芯闻· 2025-12-02 10:18
上季(2025年7-9月)村田整体接获的订单额为4,870亿日圆、较去年同期大增14.1%,其中电容订单 额大增18.1%至2,418亿日圆。 全球积层陶瓷电容(MLCC)龙头厂村田制作所(Murata Mfg)预估AI伺服器用MLCC将以年平均成长 率30%的速度呈现扩大、2030年度需求将扩增至2025年度的逾3倍水准。 共同通信报导,村田制作所社长中岛规巨1日在横滨市举行的事业战略说明会上表示,2030年度AI 伺服器用MLCC需求预估将成长至2025年度的3.3倍。随着AI伺服器处理能力提升、搭载的MLCC 数量也随之增加,村田制作所此次公布的预估值较去年11月公布的预估值进行上修。 中岛规巨指出,关于AI伺服器用MLCC需求、「是以预估年平均成长率30%左右、来作准备」。 截至日本股市2日早盘收盘(台北时间上午10点30分)为止,村田制作所涨0.40%、暂收3,237日圆, 今年迄今股价累计大涨约26%。 村田制作所10月31日公布财报资料指出,因AI伺服器及周边机器搭载的电子零件数量增加,带动 该公司产品需求扬升,今年度(2025年度、2025年4月-2026年3月)合并营收目标自1.64兆日圆上 ...
台积电2nm,疯狂扩产
半导体芯闻· 2025-12-02 10:18
Group 1 - The core viewpoint of the article is that TSMC is accelerating its capacity expansion for 2nm process technology to solidify its market dominance in advanced semiconductor foundry services, with plans to increase monthly production capacity significantly by 2027 [1][2]. - TSMC plans to raise its 2nm wafer capacity to approximately 150,000 wafers per month by 2027, compared to its current capacity of about 40,000 wafers per month, with an interim target of 80,000 to 90,000 wafers per month next year [1]. - UBS forecasts that while the 2nm process will account for less than 10% of TSMC's sales next year, this figure is expected to grow to 15-20% by 2027 [1]. Group 2 - TSMC holds a dominant position in the advanced semiconductor market, with a market share of 70.2% in the foundry sector as of Q2 this year, producing chips for major companies like Apple, Qualcomm, Nvidia, Google, and AMD [2]. - TSMC is planning to build three new 2nm fabrication plants in Taiwan to meet the increasing demand for AI chips, with a total investment of approximately 900 billion New Taiwan Dollars (around 4.21 trillion Korean Won) [2]. - Competitors like Samsung and Intel are also ramping up production to catch up with TSMC, with Samsung expected to double its wafer capacity to 21,000 wafers next year and Intel launching its 18A (1.8nm) process at its Arizona facility [3].
DDR已被HBM超越
半导体芯闻· 2025-12-02 10:18
如果您希望可以时常见面,欢迎标星收藏哦~ 全球IT巨头正持续投资人工智能,并大量收购DRAM和闪存等半导体产品。而容量相对较小的 PC内存和固态硬盘则被边缘化,面临单价上涨和供应短缺的困境。 大型PC制造商由于没有将内存和固态硬盘成本的上涨反映到产品价格中,而遭受损失。 M先生解释说:"即使生产成本只增加1美分(0.01美元,约合14韩元),电脑制造商也会反应 非常敏感。即使他们试图只降低几美分,但如果生产成本一下子增加几十美元,他们自然会蒙 受损失。" 当前内存价格上涨的驱动力主要来自人工智能相关企业。各大公司已宣布计划在未来三年内每 年将产能和处理能力翻一番。主要供应商也正集中精力生产高带宽内存(HBM)。 由于内存供应价格自10月份开始上涨,个人电脑制造商的销售成本已大幅增加,导致亏损。然 而,为了反映价格上涨的影响,明年发布的新产品预计价格将比今年至少上涨20%。 5 月下旬至 6 月初,三星电子、SK 海力士和美光等主要 DRAM 供应商通知 PC 制造商停止生 产 DDR4 产品,并宣布提高现有产品的价格。 一位不愿透露姓名的海外电脑制造商采购和供应链经理 M 先生在 2 日上午与记者的视频通话中 ...
定档12月6日!集成电路产业人才论坛报名开启!
半导体芯闻· 2025-12-02 10:18
2025浦东国际人才港论坛 集成电路产业人才论坛 将于12月6日在张江科学会堂举办 敬请期待! 识别下方二维码 参与活动报名 03 产才报告, 为进一步促进集成电路国际人才交流与发展, 多维度探讨全球人才发展创新机遇与挑战,推动集 成电路产业高质量发展,由上海市浦东新区人才工 作局指导,上海张江高科技园区开发股份有限公司 主办,上海张江浩芯企业管理有限公司、复旦大学校 友总会集成电路行业分会承办的"2025浦东国际人 才港论坛 ·集成电路产业人才论坛"将于2025年12月 6日在上海张江科学会堂举行。 对集成电路领域人才 发展研究进行专业剖析 与数据解构,揭示行业动 向与成长路径。 时间地点 时间:2025年12月06日13:30-16:00(星期六) 地点:张江科学会堂(海科路1393号)二楼张江厅A LT 01 習演趋势, 关注产业人才吸引与 发展的新动向,呈现技术 成果落地应用的典范实 例,讲述具有启示意义的 创新历程。 02 精英共论, 融通未来 集结产业、学术与研 究领域的领军人物,展开 跨学科尖端科技的深入 探讨与理念交锋。 04 专项展区, 04 聚焦"芯"生态 14:15-14:30 光计算如 ...
美国投资了一家EUV光刻机公司
半导体芯闻· 2025-12-02 10:18
Core Viewpoint - The Trump administration has agreed to invest up to $150 million in xLight, a startup focused on developing advanced semiconductor manufacturing technology, as part of its efforts to support strategically important domestic industries [1][2]. Group 1: Investment and Government Support - The U.S. Department of Commerce will provide incentives to xLight, which is working on improving the extreme ultraviolet (EUV) lithography technology critical for chip manufacturing [1][2]. - This investment utilizes funds from the 2022 CHIPS and Science Act, marking the first allocation from this act during Trump's second term [2]. - The agreement is still preliminary and subject to change, indicating that final terms have not yet been established [2]. Group 2: Technology and Innovation - xLight aims to build large "free electron lasers" powered by particle accelerators to provide more powerful and precise light sources for chip manufacturing [2][3]. - The company’s technology could potentially improve wafer processing efficiency by 30% to 40% and reduce energy consumption compared to current light sources [3]. - If successful, xLight's advancements could significantly enhance the economic viability of existing EUV lithography technology and lay the groundwork for future developments in the field [4]. Group 3: Leadership and Vision - Pat Gelsinger, former CEO of Intel, is now the executive chairman of xLight and views this venture as a significant opportunity to revive his career [1][3]. - Gelsinger has expressed a commitment to "awaken" Moore's Law, which predicts that the number of transistors on a chip will double approximately every two years [3]. - The startup has raised $40 million from investors, including Playground Global, where Gelsinger is a general partner [3].
人工智能时代,EDA巨变
半导体芯闻· 2025-12-02 10:18
如果您希望可以时常见面,欢迎标星收藏哦~ 但如上所述,AI 的到来,让更强算力的芯片成为迫切需求。在这种环境,Chiplet模式就成为自 然选择。 "通过将不同功能模块以'芯粒'形式拆分,并借助先进封装完成系统级集成,一方面可以显著提升 良率与灵活性,另一方面也为多工艺异构集成开辟现实路径。"代文亮介绍说。在他看来,这种 拆分的模式能够给芯片性能的提升带来巨大利好,先进封装的发展比我们想象的快很多,随着先 进封装持续演进,其对系统性能的影响已经不亚于制程节点。 在日前的ICCAD 2025峰会演讲中,芯和半导体科技(上海)股份有限公司(以下简称"芯和") 创始人兼总裁代文亮博士直言:"AI正在引领的是第四代工业革命,AI在云、管、端各个领域都 有渗透,且应用场景越来越多。但对于AI而言,底层的AI算力才是决定其走势的关键。" 众所周知,随着模型越来越大,AI对芯片算力的要求大幅攀升,叠加当前摩尔定律放缓带来的影 响,传统的单芯片SoC的设计模式已经不再满足终端的需求。于是,为了实现芯片设计的设计范 式跃迁,芯片的整条供应链也正在发生新的转变。 芯和正从EDA方面给行业提供新的参考。 芯片设计变了 "半导体产业 ...
这颗不被看好的芯片,终于翻身?
半导体芯闻· 2025-12-01 10:29
Core Insights - Google's TPU has gained significant attention recently, with Meta considering a multi-billion dollar contract to deploy TPUs in its data centers starting in 2027, leading to a surge in Google's stock price and a drop in NVIDIA's stock [1][20] - The TPU has evolved from a project initially deemed unpromising to a strategic asset that could challenge NVIDIA's dominance in the AI chip market [1][27] - The TPU's development has been marked by rapid iterations, with the latest version, TPU v7 Ironwood, achieving peak performance that surpasses NVIDIA's offerings [16][18] Development History - In 2013, Google faced a computational crisis, predicting that the demand for voice recognition would exceed its data center capacity, prompting the decision to develop its own ASIC chips instead of relying on NVIDIA GPUs [2][3] - The TPU project was initiated, and within 15 months, the first TPU was deployed, achieving significant performance and efficiency improvements over existing solutions [3][5] - The TPU's architecture, particularly the "Systolic Array" design, has been a key innovation, allowing for high data reuse and reduced energy consumption [5][6] Iterative Breakthroughs - TPU v2, released in 2017, marked a shift from inference to training capabilities, introducing the bfloat16 format and significantly enhancing performance for large models [9][10] - TPU v3, launched in 2018, adopted liquid cooling to manage increased power density, establishing a new standard for AI data centers [11][12] - TPU v4 introduced optical circuit switching technology, allowing for dynamic network configurations to optimize performance for varying tasks [13][14] - TPU v5p, released in 2023, aimed to balance training and inference capabilities, expanding its application scope [14] Market Position and Strategy - Google is now actively commercializing TPU, engaging with cloud service providers and large enterprises to deploy TPUs in their data centers, potentially generating billions in revenue [20][21] - The TPU's success has prompted a talent exodus from Google, with former TPU engineers founding new companies and developing competitive chips [25][26] - The competition between TPU and NVIDIA's GPUs is intensifying, with both technologies expected to coexist in a hybrid deployment model, leveraging their respective strengths [22][28] Future Outlook - The rise of TPU signifies a shift in the AI infrastructure landscape, moving towards a model that integrates cloud services with specialized chips, potentially disrupting NVIDIA's long-standing market dominance [27][28]
从EDA For AI,到EDA+AI,重构智能设计的未来
半导体芯闻· 2025-12-01 10:29
我们正置身于一个由人工智能驱动的全球性沸腾时代:从感知式 AI → 生成式 AI → 代理式 AI → 物理 AI,AI正以前所未有的速度,推动着算力体系与设计范式的更迭。每一步,都伴 随着计算量的"指数级增长"和对算力更高维度的需求。 事实上,当整个行业在向下一个 AI 阶段迈进,设计本身也成了巨大的挑战。在近日举办的 中国集成电路设计年会 ICCAD2025上,我们欣慰的看到,新的挑战正迎来创新的解法。与 加速计算一道,EDA正开启AI时代的篇章。 AI 时代主角,从"芯片—系统" 新一轮AI浪潮下,聚光灯下的主角已从"芯片"转向了"系统"。作为算力提升的核心手段之一, Scale up 范 式 正 从 单 芯 片 转 向 以 机 柜 为 单 位 的 整 体 性 能 跃 迁 。 在 这 一 架 构 下 , 几 十 至 上 百 个 GPU、CPU、内存、存储等单元,通过高速互联网络在机柜内整合为一个深度耦合、协同工作 的"超级计算单元"。 在此基础上,"超级计算单元"将作为数据中心的基本单元,通过机柜间互联,构建成更大规模的算 力集群(Scale out)。同时,算力资源池化为统一的资源空间,支持积木式按需 ...