为突破现有瓶颈，唐建石团队将芯片算力拆解为"晶体管集成密度 × 芯片面积 × 单个晶体管算力" 三个 核心要素，针对每个要素展开技术探索。 来源：半导体产业纵横 2025 年9月24日，清华大学集成电路学院副院长、长聘副教授唐建石在2025 IC WORLD 高峰论坛上， 发表题为《高算力芯片发展路径探索与存算一体芯片》的演讲。演讲围绕学院近年在高算力芯片与存算 一体芯片领域的思考、探索及实践展开，系统阐述了行业现状、技术突破与未来规划。 从他的演讲中，我们获取了以下关键信息： 唐建石指出，当前人工智能领域对算力的需求呈爆发式增长，国家计算力指数与数字经济、GDP 增长 紧密相关。其中，中国智能算力规模 2025 年已突破数十万亿亿次，且 AI 算力需求每不到六个月便实 现翻倍，这一增速远超摩尔定律驱动的硬件算力提升速度，构建更强力的芯片算力底座成为行业迫切需 求。 同时，计算芯片与存储芯片存在显著差异：存储芯片拥有统一的标准接口与定义，而计算芯片需依赖指 令集、工具链、操作系统构成的完整生态支撑。从行业格局看，美国长期主导计算芯片体系，我国则面 临双重硬件制约：一是摩尔定律逐步放缓，晶体管尺寸微缩难度加大， ...

Core Viewpoint - MediaTek has launched its flagship 5G AI chip, Dimensity 9500, which significantly enhances on-device AI capabilities, moving from experimentation to practical applications [2][12]. Group 1: AI Capabilities and Performance - The Dimensity 9500 can process long texts up to 128K characters in just two seconds, summarizing meeting notes and correcting typos automatically [3]. - Image generation on mobile devices has improved, with the Dimensity 9500 producing detailed images in 10 seconds, compared to 30 seconds with previous models [7]. - The chip supports 4K quality image generation, allowing users to create images from simple prompts in under 10 seconds [9]. - The AI applications running on the Dimensity 9500 are designed for real-world scenarios, operating locally without cloud data uploads, and consuming 50% less power than its predecessor, the Dimensity 9400 [11]. Group 2: Technological Advancements - The Dimensity 9500 features a new architecture built on a 3nm process, integrating over 30 billion transistors, resulting in a 111% increase in NPU peak performance while reducing power consumption by 56% [18][22]. - It achieved a score of 15015 on the AI Benchmark platform, nearly doubling the performance of the previous generation [19]. - The chip employs a dual NPU architecture, enhancing both performance and efficiency, and introduces a new BitNet 1.58-bit quantization framework, reducing power consumption by 50% compared to the previous model [25][28]. Group 3: Developer Support and Ecosystem - MediaTek has introduced the Dimensity AI development kit, which supports key technologies for AI model development, enabling the execution of 7 billion parameter AI models on-device [30][33]. - The company is focused on providing a standardized AI development paradigm, which is expanding the ecosystem for native AI applications [33]. Group 4: Industry Trends and Future Outlook - Major smartphone manufacturers like vivo and OPPO are set to launch devices powered by the Dimensity 9500, showcasing a shift towards advanced AI capabilities in mobile technology [36]. - The upcoming devices will feature personalized AI functionalities and enhanced performance for complex tasks, indicating a trend towards more intelligent and responsive mobile devices [39][40].

Core Viewpoint - The article discusses the challenges faced by traditional chip architectures due to the rise of generative AI models and the emergence of in-memory computing technology, which significantly enhances AI computing efficiency and is seen as a disruptive technology in the post-Moore era [1][3]. Group 1: In-Memory Computing Technology - In-memory computing technology has gained traction as it addresses the "storage wall" and "power wall" issues inherent in the von Neumann architecture, leading to a potential efficiency improvement of several times in AI computing [1][3]. - The in-memory computing chips developed by Zhichun Technology have already served over 30 clients in commercial applications, showcasing the technology's practical viability [5]. Group 2: Talent Acquisition and Development - Zhichun Technology has launched the "Genius Doctor Program" for 2026, aiming to attract top talent in semiconductor devices, circuit design, and AI algorithms, reflecting the industry's talent competition amid rapid technological advancements [1][7]. - The program offers a unique growth system that includes mentorship and rotation across core R&D positions, allowing participants to gain comprehensive experience in the technology development process [7][10]. Group 3: Industry Trends and Future Outlook - The semiconductor industry is expected to face a talent shortage of over 300,000 professionals by 2025, highlighting the urgency for companies to develop and attract skilled individuals [1]. - The current phase of in-memory computing technology is critical as it transitions from "production validation" to "scale application," indicating a pivotal moment for the industry [12].

Core Insights - The article highlights the strong growth in orders for advanced logic semiconductor equipment driven by AI chips, with two major industry leaders experiencing a 40% year-on-year increase in orders in the first half of the year [1] - NVIDIA's launch of the Rubin CPX is expected to significantly lower token generation costs, potentially stimulating overall demand for AI applications as this product is anticipated to grow alongside the overall increase in AI workloads [1] - The rise of AI terminals may disrupt the traditional separation of "computation" and "storage" architectures, with "compute-storage integration" or "near-storage computing" likely to come to the forefront, driving demand for corresponding equipment and materials [1]

Core Viewpoint - The 2025 Global AI Chip Summit will be held on September 17 in Shanghai, focusing on the theme "AI Infrastructure, Smart Chip New World," addressing the new infrastructure wave in the AI era and the breakthroughs in China's chip industry under large models [2][3]. Group 1: Event Overview - The summit will feature over 180 industry experts sharing insights on cutting-edge research, innovations, and industry trends, making it a significant platform for understanding AI chip developments [2]. - The event will consist of a main forum, specialized forums, technical seminars, and an exhibition area, providing a comprehensive agenda for attendees [2][3][5]. Group 2: Main Forum Highlights - The opening report will be delivered by Professor Wang Zhongfeng, focusing on "Shaping the Intelligent Future: Architectural Innovation and Paradigm Shift of AI Chips," discussing solutions to overcome bottlenecks in AI chip development [7]. - Key speakers include leaders from major companies such as Huawei and Yuntian Lifei, discussing trends in AI development and the strategic positioning of AI chips [7][8][9]. Group 3: Specialized Forums - The Large Model AI Chip Specialized Forum will address the competitive landscape of large models and the infrastructure needed for AI, emphasizing cost-effectiveness as a critical factor [18][19]. - The AI Chip Architecture Innovation Forum will explore new chip architectures, including wafer-level chips and RISC-V based solutions, highlighting the need for innovative approaches in the face of technological constraints [22][24]. Group 4: Technical Workshops - The workshops will focus on topics such as memory wall issues in traditional architectures and the importance of storage-computing integration in AI chip design [32][33]. - Experts will discuss advancements in DRAM near-memory computing architectures and the challenges of integrating heterogeneous systems for AI applications [34][35]. Group 5: Exhibition Area - The exhibition will feature over 10 exhibitors, including leading companies like Achronix and Sunrise, showcasing their latest technologies and solutions in the AI chip sector [3].

Group 1 - The company, Yihualu (300212), is focusing on enhancing its competitiveness and achieving performance improvement in areas such as smart transportation, data elements, and integrated computing and storage [1]

Core Insights - The Chinese chip industry is facing three major challenges: limitations in advanced process technology, reliance on a closed software ecosystem, and bandwidth bottlenecks due to traditional architecture [1][2][3][4] Group 1: Challenges in the Domestic Chip Industry - The lack of advanced manufacturing processes has resulted in a bottleneck in computing density, with current domestic 3nm/5nm technologies still in the R&D phase and unable to meet the demands of large AI models [2] - The domestic AI chip industry is heavily dependent on Western closed-source ecosystems, particularly the CUDA ecosystem, which monopolizes AI model training and inference software, leading to a situation where high-performance chips may lack compatible software [3] - Traditional von Neumann architecture separates computing and storage units, causing data to be frequently moved via buses, creating a "memory wall" bottleneck that significantly reduces inference efficiency as model parameters scale to hundreds of billions [4] Group 2: 3D-CIM Technology as a Solution - The 3D-CIM (3D Compute-in-Memory) technology, introduced by Micronano Core, integrates computing capabilities within storage, addressing the exponential growth in computing demands for AI models [5] - This technology utilizes SRAM compute-in-memory combined with DRAM 3D stacking to perform computations within the memory, fundamentally eliminating data transfer overhead and is seen as a key path for sustaining computing growth in the post-Moore's Law era [5] - The core breakthrough of 3D-CIM lies in its SRAM compute-in-memory design, which allows for in-situ tensor computations, significantly enhancing computing density and achieving performance comparable to traditional NPU/GPU at a lower manufacturing cost [5][6] Group 3: Ecosystem and Application Prospects - The open and flexible RISC-V architecture complements the 3D-CIM technology, meeting the high parallelism and low power consumption needs of AI models while alleviating external process restrictions [7] - Micronano Core is collaborating with upstream and downstream enterprises to promote the ecological implementation of 3D-CIM technology and RISC-V architecture [8] - The application prospects for 3D-CIM technology are categorized into short-term, mid-term, and long-term, with initial applications in edge AI devices, followed by cloud-based AI model applications, and eventually expanding into embodied intelligence applications [8]

Group 1 - The technology sector in A-shares continues to perform well, with a focus on systematic investment strategies in the technology field [1] - The underlying hardware systems are crucial for advancements in AI and autonomous driving, with computing power being likened to "oil" in the digital age [3] - China still relies on imports for high-end chips, advanced manufacturing equipment, and key materials, but technological breakthroughs and policy support are accelerating the restructuring of the semiconductor industry [6] Group 2 - The concept of "storage-compute integration" is emerging, which allows for data processing at the storage unit level, addressing the challenges of power consumption and latency in traditional computing architectures [9] - Advanced packaging technology, particularly optical-electrical co-packaging, is becoming essential for improving data transmission efficiency and reducing losses, with significant market growth potential [13] - The focus on semiconductor leaders and technology innovation is critical for capturing investment opportunities in the tech sector [14]

Group 1 - The technology sector in A-shares continues to perform well, with a focus on systematic investment strategies in the technology field [1] - The underlying hardware systems are crucial for advancements in AI and autonomous driving, with computing power being likened to "oil" in the digital age [3] - China still relies on imports for high-end chips, advanced manufacturing equipment, and key materials, but technological breakthroughs and policy support are accelerating the restructuring of the semiconductor industry [6] Group 2 - The concept of "storage-compute integration" is emerging, which allows for data processing at the storage unit level, addressing the challenges of power consumption and latency in traditional computing architectures [9] - Advanced packaging technology, particularly optical-electrical co-packaging, is becoming essential for improving data transmission efficiency and reducing losses, with significant market growth potential [14] - The focus on semiconductor leaders is critical for capturing opportunities in technological innovation [15]

Core Viewpoint - Alibaba is developing a new generation of AI chips focused on multifunctional inference scenarios, aiming to fill the market gap left by NVIDIA's H20 exit [1][3]. Chip Development and Specifications - The new chip utilizes domestic 14nm or more advanced processes, supported by local foundries like Yangtze Memory Technologies, integrating high-density computing units and large-capacity memory with an expected LPDDR5X bandwidth exceeding 1TB/s, targeting a single-card computing power of 300-400 TOPS (INT8), comparable to H20's approximately 300 TOPS [1][3]. - Compared to NVIDIA's H20, Alibaba's chip offers full-scene compatibility, supporting FP8/FP16 mixed precision computing and seamless integration with the CUDA ecosystem, reducing migration costs by over 70% [3]. - Alibaba has urgently increased its order for the Cambricon Siyuan 370 chip to 150,000 units, which is based on a 7nm process and utilizes Chiplet technology, integrating 39 billion transistors and achieving a measured computing power of 300 TOPS (INT8) with a 40% improvement in energy efficiency [5]. Market Strategy and Production Capacity - The Cambricon Siyuan 370 chip is expected to cover 60% of Alibaba Cloud's inference demand by Q2 2025 and supports multi-card interconnection via PCIe 5.0, facilitating user growth for Tongyi Qianwen [5]. - Alibaba collaborates with Yangtze Memory Technologies to develop AI chips focusing on overcoming storage bottlenecks, achieving a storage density of 20GB/mm² and read/write speeds of 7000MB/s, a 40% improvement over the previous generation, expanding local storage capacity to 128GB [5][6]. - To ensure mass production, Alibaba employs a dual-foundry backup strategy, with SMIC's 14nm production line handling basic chip production, achieving a stable yield of over 95% and a monthly capacity of 50,000 units [6]. Future Roadmap - Alibaba's three-step strategy includes: - Short-term (2025-2026): Focus on 7nm/14nm inference chips to quickly capture market share through ecosystem compatibility [10]. - Mid-term (2027-2028): Launch 4nm training chips targeting a computing power of 1 EFLOPS, competing with NVIDIA's H100 [10]. - Long-term (post-2030): Explore disruptive technologies like photonic computing and integrated storage-computing solutions, with the first commercial photonic AI chip already released, promising a speed increase of 1000 times and a 90% reduction in power consumption compared to GPUs [10]. - Alibaba's path to domestic computing power is characterized as a dual battle of technological breakthroughs and ecosystem reconstruction, aiming to disrupt NVIDIA's monopoly through a "compatibility-replacement-surpassing" strategy [10][11].