昨天晚上，AMD CEO 苏姿丰，又在华尔街掀起了一阵风；她说：到 2030 年，AI 数据中心市场将突破 1 万亿美元。 这句话听起来像在吹牛，但，如果你了解过去 10 年她如何让AMD起死回生，就知道，这个人从来不是说空话的人。 01 首先，为了降低认识门槛，快速过下她是谁？ 很多人知道英伟达的黄仁勋，因为他特别会造梦；但不少人对苏姿丰这个名字有点陌生。她的英文名叫 Lisa Su，是AMD现任的CEO兼董事长。 2014 年，她从一个纯技术背景的工程师，一步步走到这家老牌芯片公司的最高位。 要说履历，她堪称「理工界的女天花板」： 麻省理工学院（MIT）电气工程博士，先后在 IBM 和 Freescale 半导体担任高管；她的专长不是营销、演讲，是让晶体管如何变得更聪明。 接手 AMD 那一年，公司基本快要崩盘，市场份额被英特尔吊打，市值不到 30 亿美元，连员工都在担心明天还有没有班上。 结果不到 10 年，她硬是把 AMD 拉回了牌桌。CPU 端推出的 Ryzen 系列，一口气击穿了 Intel 的防线；GPU 端的 Instinct 系列，对标的正是 NVIDIA。 如果你对这些产品不太熟，也没关系 ...

Core Viewpoint - AMD is planning to adopt multi-chip module (MCM) architecture for consumer-grade GPUs, leveraging its experience from the Instinct MI200 AI accelerator series, which was the first to implement MCM design [3][5]. Summary by Sections Multi-Chip Module (MCM) Concept - The MCM concept is gaining traction in the graphics processing industry due to limitations of single-chip designs, with AMD being a key player in this transition [3]. - AMD's Instinct MI200 series integrated multiple chiplets, including graphics processing cores, HBM stacked memory, and I/O chips [3]. Technical Innovations - AMD's new patent reveals a "smart switch data structure circuit" that connects compute chiplets with memory controllers, optimizing memory access and reducing latency to nanoseconds [4]. - Each Graphics Compute Die (GCD) will feature L1 and L2 caches, similar to AI accelerators, and will have access to a shared L3 cache through the switch, minimizing global memory access [4]. Ecosystem and Competitive Advantage - AMD has a complete ecosystem in place, utilizing TSMC's InFO-RDL bridging technology and a specific version of Infinity Fabric for chip interconnects [5]. - The integration of gaming and AI architectures into a unified UDNA architecture positions AMD favorably against competitors, despite the complexities associated with chiplet designs [5]. Future Outlook - AMD aims to address latency issues experienced in previous architectures, such as RDNA 3, through innovative solutions like the shared L3 cache [5]. - The market may not see the full impact of these innovations until the release of UDNA 5 [5].

Core Insights - AMD's Instinct MI350 series utilizes HBM products from Samsung Electronics and Micron [1] - The series features 8 units of 36GB HBM3E memory, with a 12Hi stacking configuration and 8Gbps performance [1] - The total memory capacity is 288GB, with a bandwidth of 8TB/s [1]

Core Viewpoint - AMD reported strong Q1 2025 earnings, with significant year-over-year growth in revenue and net income, despite a slight quarter-over-quarter decline in revenue [2][5][10]. Financial Performance - Q1 2025 revenue reached $7.438 billion, a 36% increase from $5.473 billion in Q1 2024, but a 3% decrease from $7.658 billion in Q4 2024 [3][10]. - Operating income surged to $806 million, up 2139% year-over-year, while net income rose to $709 million, reflecting a 476% increase [2][3]. - Non-GAAP gross margin was reported at 54%, with diluted earnings per share at $0.96 [2][10]. Business Segments - Revenue sources include data center, client, gaming, and embedded segments [7]. - Data center revenue was $3.7 billion, up 57% year-over-year, driven by strong sales of AMD EPYC CPUs and Instinct GPUs [10]. - Client segment revenue reached $2.9 billion, a 28% increase, with client business growing 68% due to demand for the latest "Zen 5" architecture [10]. AI Chip Development - AMD began providing samples of the AI chip Instinct MI350 series to multiple customers, with production expected to ramp up in the second half of the year [5][6]. - The MI350 series offers 35 times the throughput and performance compared to the MI300X [5]. Regulatory Impact - Recent U.S. export controls on the MI308X chip to China are expected to impact AMD's revenue by approximately $1.5 billion for the full year 2025 [5][10]. - The company anticipates Q2 2025 revenue to be around $7.4 billion, with a potential $800 million impact from the new export controls [10]. Market Outlook - AMD's CEO expressed confidence in achieving double-digit revenue growth for 2025, supported by a strong product portfolio and increasing market share in AI workloads [6][10]. - The company completed the acquisition of ZT Systems to enhance its capabilities in the data center AI accelerator market, which is projected to reach $500 billion by 2028 [8].