以下文章来源于傅里叶的猫 ，作者小小 傅里叶的猫 . 芯片EDA大厂资深工程师，半导体AI行业解读及研报分享 申请入围"中国IC独角兽" 是说芯语，欢迎关注分享 美国或设 GPU 内存带宽上限 1.7-1.8TB/s，若如此英伟达 H20 可能弃用 HBM 内存改用 GDDR6，降级 版 H20 性能仍可能强于使用 GDDR6 的游戏 GPU，如RTX5090D。 今天又看到另一篇研报，是广发海外的一个分析，认为该 GPU可能命名为 6000D，即 B40，极有可能 在 7 月初发布，其搭载的 GDDR7 速率约为 1.7TB/s（对比 H20 的 4TB/s）。在 NVLink 单向速度约 550GB/s 及 CUDA 的支持下，广发预计到 2025 年底出货量将达约 100 万片。 对于很多没有公布的数据，我们都要结合多家的报告来看，因为每家的分析可能多多少少有些不同的地 方。就像我们上篇文章中提到的华为昇腾系列2025年出货量调研的文章，几家机构给出的数据都是在 70w~85w之间，而在IDC的一个分析中，提到2024年华为昇腾的出货量已经达到了64万，那按正常的增 长速度，加上H20被禁，今年绝对不会 ...

申请入围"中国IC独角兽" 半导体高质量发展创新成果征集 图片来源：中国商务部 2025年5月21日，中国商务部发布一份声明警告，针对 任何组织和个人 执行或协助执行美国对中国先进 计算芯片（包括科技巨头华为的产品）的限制措施，中国将会对其采取法律行动。 中方注意到， " 美国商务部近日发布指南，以所谓 推定违反 美国出口管制为由，企图在全球禁用中国 先进计算芯片，包括特定的华为昇腾芯片。" 中方认为， 美方措施是典型的 单边霸凌和保护主义 做法，严重损害全球半导体产业链供应链稳定，剥 夺其他国家发展先进计算芯片和人工智能等高科技产业的权利。这违反国际法和国际关系基本准则 ， 严重损害中国企业正当权益，危害中国发展利益。 中方强调， 美方措施涉嫌构成对中国企业采取的 歧视性限制措施 。任何组织和个人 执行或协助执行 美方措施，将涉嫌违反 《中华人民共和国反外国制裁法》 等法律法规，须承担相应法律责任。 今天的这一声明呼应了中国商务部在5月15日发出的 另一则警告 ：中国将针对美国的芯片管制采取"坚 决措施"。 中国商务部发言人在5月19日还表示，美国针对华为高端芯片的措施严重破坏 中美日内瓦会谈共识 。5 月 ...

半导体高质量发展创新成果征集 申请入围"中国IC独角兽" 马斯克的人工智能公司xAI正在美国田纳西州孟菲斯郊外建造一个拥有100万颗GPU的巨型工厂 Colossus，并将该地区打造成为"高科技制造中心"。 2025.05. 21 本文字数：1167，阅读时长大约2分钟 作者 | 第一财经 钱童心 当地时间5月20日，特斯拉CEO埃隆·马斯克在多哈参加卡塔尔经济论坛时表示，他希望在未来五年 内继续领导特斯拉。 他最近还表示，特斯拉和xAI正在向英伟达、AMD等公司采购更多芯片。xAI计划建一个拥有100万 颗GPU的大型工厂。 马斯克发表这些言论之际，特斯拉正在面临投资者的质疑。尽管近一个月特斯拉股价上涨超过 40%，但市值仍未收复1万亿美元。截至5月20日收盘，特斯拉今年股价累计下跌近15%。 今年4月，特斯拉公布财报显示，该公司今年第一季度汽车业务收入下降20%，净利润下降71%。 此前，马斯克领导了美国总统特朗普设立的"政府效率部门"，但大刀阔斧地精简美国公职部门人员 的举动也引发巨大争议。对此，马斯克表示，未来计划大幅削减政治竞选的支出。 马斯克 暗示 他将把精力更多地聚焦在自己公司的治理。在担任特斯 ...

Core Viewpoint - The article highlights the impressive performance of Xiaomi's new SoC, the "玄戒" (Xuanjie), which has managed to compete with MediaTek's long-standing expertise in SoC performance and power management within a short span of four years [2][3]. Group 1: SoC Performance Comparison - Xiaomi's Xuanjie SoC outperforms MediaTek's Dimensity 9400 in single-core performance, scoring 3100 compared to 2900, indicating a significant advancement in technology [2]. - The architecture of the Xuanjie includes a 10-core setup with high clock speeds, which contributes to its superior performance metrics [2]. - The Xuanjie SoC's design choices, such as the use of external 5G baseband chips, may lead to lower power density and better thermal management compared to integrated solutions [2]. Group 2: Development and Investment - Xiaomi has invested a total of 135 billion RMB in the development of the Xuanjie SoC over four years, with a projected future investment of 500 billion RMB over the next decade [3]. - The development team for the Xuanjie consists of approximately 2500 members, showcasing a significant commitment to building a robust in-house capability [3]. - The rapid development of the Xuanjie SoC is compared to other industry players, such as Zeku and Huawei's HiSilicon, which took much longer to achieve similar levels of performance [3]. Group 3: Market Position and Future Outlook - The introduction of the Xuanjie SoC positions Xiaomi among the top tier of the semiconductor industry, potentially achieving what Zeku could not complete [3]. - The article suggests that Xiaomi's investment strategy appears more genuine compared to other companies, indicating a strong commitment to semiconductor development [3]. - The future performance of the Xuanjie in real-world applications remains to be seen, but initial results are promising [3].

Core Viewpoint - The bankruptcy of Jiangxi Chuangcheng Microelectronics highlights the challenges faced by small and medium-sized enterprises in the semiconductor industry during a period of deep adjustment and market contraction [2][5][7] Company Overview - Jiangxi Chuangcheng Microelectronics, established in October 2015, was a national high-tech enterprise focused on digital signal processing (DSP) chip design, with a registered capital of 10 million yuan [3] - The company developed over 10 proprietary chips and held more than 150 authorized patents, serving clients like Philips, Tencent, and Lenovo [3][4] - Despite its technological achievements, the company faced severe operational difficulties, with unpaid debts exceeding 12.82 million yuan and a significant reduction in workforce [3][4] Industry Challenges - The decline of Jiangxi Chuangcheng reflects broader issues in the semiconductor sector, including a downturn in consumer electronics and increased cost pressures in the automotive electronics market [5][6] - High research and development investment (5.85% from 2016 to 2018) did not translate into commercial success, leading to cash flow issues [5] - The inability to sustain operations amid a tightening capital environment has exposed the vulnerability of smaller firms in the industry [5][6] Industry Warnings - Jiangxi Chuangcheng is not an isolated case; several chip companies have exited the market since 2019, indicating a persistent and harsh industry winter [7] - Companies must not only overcome technological barriers but also establish sustainable business models, diversify customer bases, and balance R&D investment with commercialization [7] - The semiconductor industry faces a critical challenge in creating a complete ecosystem from technology development to market collaboration in the context of rationalized capital and shrinking market demand [7]

Core Viewpoint - The semiconductor industry faces increasing supply chain disruptions, necessitating improved risk management mechanisms and digital tools for real-time monitoring and prevention of supply chain risks [1][3]. Group 1: Supply Chain Challenges - Geopolitical conflicts have triggered a "neon gas crisis," leading to increased production costs [5] - Fluctuating tariff policies force adjustments in global supply chain layouts [5] - Natural disasters have caused factory shutdowns, resulting in soaring raw material prices [5] - Market demand volatility hampers timely development of quality supplier resources [5] Group 2: Risk Management Trends - Transitioning from localized monitoring to global scanning for comprehensive risk awareness [6] - Moving from manual selection to tiered management for precise risk alerts [6] - Evolving from unilateral management to collaborative risk mitigation [11] - Shifting from fragmented analysis to panoramic decision-making for closed-loop risk management [11] - Upgrading from repetitive processes to agile system integration [11] Group 3: Technological Solutions - The launch of a supply chain risk management SaaS platform aims to assist semiconductor companies in transitioning from experience-based approaches to data-driven reasoning [3] - The platform integrates artificial intelligence and various business scenarios to enhance digital risk management across multiple industries [10]

申请入围"中国IC独角兽" 半导体高质量发展创新成果征集 今天，黄仁勋在中国台湾地区Computex 2025的主题演讲结束后，接受了 Stratechery的主理人 Ben Thompson的 一次面对面访谈。 在这次访谈中，黄仁勋比较罕见地对AI扩散规则、芯片出口管制、美国对华AI竞争的战略 等问题发表了观点。 黄仁勋认为，AI扩散规则限制其他国家使用美国技术，这个目标从一开始就错了。美国应该加速向全球推广 美国技术，否则一切都为时已晚。如果目标是让美国处于领先位置，那么AI扩散规则完全是反作用。美国公 司不能放弃中国， 如果不去中国竞争，就会允许中国生态系统在美国公司缺席的情况下发展壮大。这将导致 新平台出现，但这些平台将不是美国的。尤其在AI技术迅速向全球扩散的时候，中国的领先地位及其技术将 扩散到世界各地。 在黄仁勋看来， 美国要想赢得和中国的竞争，不能靠 打压竞争对手，而是要靠快速发展和创新。因为美国商 务部 禁止H20对华出口，英伟达 不仅损失了55亿美元（库存注销），还放弃了中国市场150亿美元的销售收 入，美国政府也损失了大约30亿美元的税收，且还威胁到CUDA平台。 任何认为仅仅通过禁止向 ...

英伟达CEO黄仁勋近期动作频繁，试图在技术管制与市场拓展间寻找平衡。5月19日，他在 Computex2025宣布 联合台积电、富士康在中国台湾建立AI超级计算机 ，并推出个人AI计算机 DGXSpark，强调"AI工厂"时代的到来。此前，他于4月访华期间重申中国市场的重要性，并透露英伟达 正扩建上海研发中心，计划引入2000名工程师，聚焦定制化芯片设计与自动驾驶技术开发。 Blackwell架构的技术革新集中体现在 B300芯片 上。作为Blackwell Ultra系列的核心，B300采用台积电 4NPT SMC工艺制造，通过10TB/s的片间互联技术将GPU裸片连接成统一计算单元，实现了1.8TB/s的双 向NVLink带宽，支持多达576块GPU的无缝协作。其FP4性能较前代提升1.5倍，配合第二代Transformer 引擎，可在4位浮点精度下实现算力与模型规模的翻倍。然而，为满足美国出口管制要求，Blackwell特 规芯片将 放弃HBM高带宽内存，改用GDDR7 ，这一调整虽降低了性能上限（预计较原版Blackwell下 降约30%），但仍显著优于被禁的H20。 对比H20的96GB HBM3内 ...

Core Viewpoint - The article emphasizes the urgent need for technological autonomy in the semiconductor industry, particularly in light of the PACMAN vulnerability that exposes significant risks in data security and reliance on ARM architecture [2][3][5]. Group 1: PACMAN Vulnerability - The PACMAN vulnerability undermines the core protective mechanisms of ARM architecture, allowing attackers to bypass security measures and gain control over systems without triggering alarms [3][6]. - The vulnerability affects widely used chips in smartphones, data centers, and industrial control systems, posing threats to personal privacy and national infrastructure [3][5]. - ARM's unique business model complicates the situation, as the majority of domestic market chips are based on ARM architecture, which is vulnerable to exploitation [5][6]. Group 2: Challenges of ARM Architecture - ARM's "black box" architecture means that hardware-level design flaws cannot be patched and require hardware replacement, increasing the risk of exploitation [6]. - The lengthy response time to vulnerabilities allows hackers to exploit weaknesses, while undisclosed debugging interfaces may serve as hidden backdoors for data theft [6]. Group 3: Solutions for Technological Autonomy - The industry must focus on three dimensions to overcome the challenges posed by ARM: 1. **Technological Autonomy**: Promote the large-scale application of domestic instruction set chips to reduce dependency on a single architecture [8]. 2. **Security Systematization**: Establish comprehensive security standards covering the entire lifecycle of chip design, manufacturing, and application [9]. 3. **Ecosystem Collaboration**: Encourage deep collaboration among instruction sets, operating systems, and application software to create an open-source ecosystem [10]. Group 4: Implications of Vulnerabilities - The PACMAN vulnerability highlights the fragility of traditional technology systems and underscores the importance of independent innovation to safeguard digital sovereignty [10].

Core Viewpoint - Nvidia is collaborating with TSMC and Foxconn to establish an AI supercomputer in Taiwan, which is expected to enhance its competitive edge in the AI sector and contribute to the development of Taiwan's technology industry [2][3]. Group 1: Nvidia's AI Supercomputer Initiative - Nvidia's CEO Jensen Huang announced the partnership with TSMC and Foxconn during the Computex 2025 keynote, highlighting the potential of this collaboration to create a powerful AI supercomputer [2]. - TSMC's advanced semiconductor manufacturing capabilities will provide high-performance chip support for the AI supercomputer, while Foxconn's expertise in electronics manufacturing will ensure efficient production and delivery [2]. - The establishment of the AI supercomputer is significant for advancing AI technology and applications, offering robust computing power to accelerate AI model training and optimization [3]. Group 2: Nvidia's Market Strategy in China - Nvidia has been actively expanding its presence in the Chinese market, with plans to expand its Shanghai campus to accommodate growth [3]. - Despite the expansion, Nvidia stated that GPU design modifications will not occur in China, indicating a strategic focus on maintaining certain operations outside the region [3]. - The AI supercomputer initiative is part of Nvidia's global AI strategy, which aims to enhance Taiwan's influence in the AI field and attract more related enterprises and talent [3].