TPU TPU 、、 LLPUPU 、、 GGPUPU AAII芯芯⽚⽚的的过过去去、、现现在在与与未来未来 历史演进 从图形处理到AI基⽯的华丽转⾝ 架构对⽐ 专⽤化与通⽤性的技术权衡 未来展望 异构计算与边缘AI的新时代 核⼼洞察 在⼈⼯智能浪潮席卷全球的今天，算⼒已成为驱动技术⾰命的核⼼引擎。在这场算⼒竞赛中，图形处理 器（GPU）、张量处理器（TPU）和语⾔处理器（LPU）等专⽤芯⽚扮演了⾄关重要的⻆⾊。 GPU凭借NVIDIA的CUDA⽣态，从图形渲染领域华丽转⾝，成为AI训练的基⽯； TPU源于⾕歌对内部算 ⼒危机的"未⾬绸缪"，以专⽤架构重塑了计算效率； LPU则由前TPU团队再创业⽽⽣，精准切⼊推理市 场，以确定性执⾏架构挑战传统范式。 这三款芯⽚的诞⽣与发展，共同谱写了AI硬件从通⽤到专⽤、从训练到推理的演进史诗，并将在未来持 续塑造AI技术的边界与格局。 1. 回顾历史：AI芯⽚的诞⽣与初⼼ 1.1 GPU：从图形处理到AI基⽯的华丽转⾝ ⻩仁勋的远⻅：CUDA⽣态的构建 在⼈⼯智能浪潮席卷全球之前，NVIDIA的核⼼业务聚焦于为电⼦游戏提供⾼性能的图形处理器。然⽽， 公司创始⼈兼CEO⻩ ...

文 | 下海fallsea，作者 | 胡不知 2025年12月24日，平安夜的硅谷没有温情。当大多数人沉浸在节日氛围中时，AI算力圈传来一则足以 改写行业格局的消息：英伟达宣布以200亿美元现金，与曾喊出"终结GPU霸权"的AI芯片初创公司Groq 达成技术许可协议。 "这不是收购，却胜似收购。"伯恩斯坦分析师Stacy Rasgon一针见血地指出，"本质是英伟达用金钱换时 间，把最危险的颠覆者变成自己人，同时规避反垄断审查的障眼法。" 这场交易的背后，是AI产业的历史性转折——从集中式模型训练，全面迈入规模化推理落地的新阶 段。推理市场正以年复合增长率65%的速度扩张，预计2025年规模突破400亿美元，2028年更是将达到 1500亿美元。而英伟达的GPU霸权，在推理赛道正遭遇前所未有的挑战：谷歌TPU凭借成本优势抢食大 客户，AMD MI300X拿下微软40亿美元订单，中国的华为昇腾在本土市场份额已飙升至28%。 曾被视为"GPU终结者"的Groq，为何最终选择与英伟达联手？200亿美元的天价交易，能否帮英伟达守 住算力王座？这场"招安"背后，更折射出AI芯片行业创新者的集体困境：当技术颠覆者撞上巨头的 ...

Core Viewpoint - RISC-V has the potential to integrate the characteristics of CPU, GPU, and AI processors, breaking through the ecological barriers of CUDA [47] Group 1: AI and Computing Power - The eighth GAIR Global AI and Robotics Conference will be held in Shenzhen, focusing on the core of intelligent systems—computing power [2] - Computing power is not just a reflection of hardware performance but a capability system to complete tasks under resource and time constraints [3] - The rapid growth of generative AI's demand for computing power necessitates heterogeneous computing (CPU + XPU) as CPUs alone cannot meet real-world needs [11][16] Group 2: Software and Ecosystem - The true determinant of computing power release is the software and application ecosystem, rather than the hardware itself [20] - The ecosystem includes all software that runs on processors, and the productivity is generated by application software, not the chips [24] - The x86 ecosystem has a significant market share and inertia, making it challenging for new architectures to compete [26] Group 3: RISC-V and Market Challenges - RISC-V's openness presents new possibilities, but openness does not guarantee success; many open CPUs have failed commercially [27][28] - RISC-V faces commercialization difficulties, particularly in complex computing fields, due to an immature software ecosystem [29] - The need for a robust software ecosystem is critical for RISC-V to succeed in the competitive landscape [20][29] Group 4: Future Directions - The future of computing architecture may return to a CPU-centric model, with RISC-V having the potential to unify CPU, GPU, and AI processor characteristics [47] - The importance of building a domestic computing ecosystem is recognized at the national level to avoid dependency on foreign technologies [33] - Successful chip development hinges on the ability to create a comprehensive software ecosystem that adds significant value to products and services [34][45]

公众号记得加星标⭐️，第一时间看推送不会错过。 在最近一次RISC-V 峰会上，谷歌数据中心性能工程总监 Martin Dixon 发表了一场精彩的演讲。他 以谷歌成功过渡到基于 ARM 的服务器为例，带领听众进行了一次"公路旅行"，探讨了谷歌将 RISC- V 集成到其庞大的仓库级计算基础设施中的愿景。Dixon 概述了将 RISC-V 应用于数据中心规模所 面临的机遇、挑战和必要条件。 谷歌的异构计算之旅始于其基于通用 x86 平台的早期阶段，并在不断变化的需求中迎来了 27 周年纪 念。2010 年代中期，该公司开始尝试 ARM 架构，并遵循了 2014 年发布的 ARM 服务器规范。这 促成了 2022 年 Tau T2A ARM 实例的推出，以及最近推出的定制 Axion ARM 处理器。如今，谷歌 的数据中心已经混合部署了 x86、ARM 和新兴架构，包括早期的 RISC-V 组件。Dixon强调，异构 性和专业化对于克服摩尔定律放缓至关重要，能够实现规模化更高的效率和性能。 RISC-V 的 开 放 性 和 定 制 潜 力 令 人 兴 奋 ， 但 Dixon 也 警 告 说 ， 如 果 没 有 ...

Core Viewpoint - The article discusses the evolution of computing power as a fundamental infrastructure and explores the necessary technological paths, ecological strategies, and business logic to navigate through cycles and occupy the top of the future value chain [1][3]. Group 1: Current State and Future of Computing Power - The GAIR 2025 conference focuses on the core of intelligent systems—computing power, examining its architecture, ecosystem, tools, and industrialization for the next decade [2]. - The conference features discussions on the current state and future of domestic computing power, emphasizing the need for a unified approach to overcome existing challenges [6][10]. Group 2: Key Insights from Experts - Tang Zhimin, a prominent figure in the microelectronics field, emphasizes the importance of software-defined computing power to break through chip technology barriers and highlights the critical role of software ecology in the computing chip industry [4][6][8]. - Liu Fangming discusses the challenges faced by domestic large models, advocating for a shift from "barbaric growth" to a more systematic and open ecosystem [10][12]. - Li Xingyu from Suiruan Technology points out that the domestic computing power industry is entering a phase of elimination, where software ecology will be a key determinant of success [14][16]. Group 3: Technological Innovations and Trends - Wang Hua from Moore Threads highlights the necessity of large-scale clusters for training large models, presenting data that shows significant reductions in training time with increased cluster size [19][21]. - Luo Yi from Yuntian Lifei predicts a pivotal shift in AI chip consumption from training to inference by 2025, driven by the explosive demand for inference capabilities [25][27]. - Zhao Zhanxiang from IO Capital discusses the need for diverse technological paths in the face of export controls, emphasizing the importance of system-level architecture and process innovation [30][32]. Group 4: Future Directions and Industry Consensus - The article concludes with a call for continued exploration and innovation in the computing power ecosystem, emphasizing the importance of collaboration among academia, industry, and investment sectors to shape the future landscape [35][39].

更为深层的技术矛盾，在于处理器微架构的设计哲学。现代CPU普遍依赖"推测执行"技术，通过分支 预测来提前执行指令以保持流水线满载，这种机制在处理逻辑复杂的通用程序时表现优异。然而，AI 和机器学习工作负载主要由大规模的向量和矩阵运算构成，且内存访问模式往往呈现出高度的不规则 性。在这种场景下，推测执行容易出现预测失败，导致流水线频繁刷新。被丢弃的计算指令不仅未能产 生有效产出，反而造成了额外的能源浪费与延迟。 针对通用架构在AI负载下的局限性，处理器行业正在经历第一层维度的革新：微架构层面的去推测 化。近期获得美国专利商标局专利认证的"基于时间的确定性执行模型"代表了一种新的设计思路。该模 型摒弃了复杂的推测机制，引入带有时间计数器的向量协处理器，采用静态调度策略。在这一架构下， 指令仅在数据依赖关系完全解决且操作数就绪的确定时刻，才会被分发至执行单元。 由于执行顺序和时间是预先规划且确定的，芯片设计可以省去复杂的寄存器重命名和乱序执行控制逻 辑，从而在矩阵计算等任务中以更低的晶体管开销和功耗实现高可扩展性。这种确定性执行模型在保持 与RISC-V等标准指令集兼容的同时，从底层逻辑上适配了AI计算对高吞吐量和 ...

Core Insights - The collaboration between NVIDIA and Synopsys aims to integrate advanced computing technologies, including AI-assisted engineering and digital twin platforms, to enhance Synopsys' product offerings and accelerate market strategies [2][11] - NVIDIA's $2 billion investment in Synopsys at a price of $414.79 per share signifies a long-term commitment to this partnership, which is expected to reshape the engineering simulation landscape [1][11] Group 1: Collaboration Details - The partnership will leverage NVIDIA's GPU technology to enhance Synopsys' EDA, simulation, and multiphysics product lines, moving beyond traditional CPU dominance in chip design [1][2] - Synopsys plans to utilize NVIDIA's tools to accelerate various engineering processes, including chip design, physical verification, and optical simulation [2][3] - The collaboration is characterized by its broad scope, aiming to integrate multiple engineering phases from transistor-level design to final physical products [2][11] Group 2: Technical Aspects - Both companies acknowledge that while some workloads currently utilize GPUs, significant algorithmic restructuring is necessary to fully capitalize on GPU acceleration [4][5] - The transition to GPU-accelerated workflows is expected to be gradual, potentially extending into 2026 and 2027, as deeper structural changes are required for multiphysics and electromagnetic workflows [5][7] - The focus on AI integration is crucial, as it will enhance Synopsys' AI technology stack and improve applications in solvers, simulators, and digital twins [7][19] Group 3: Market Opportunities - The collaboration is seen as a way to expand the simulation and modeling market by lowering costs and speeding up processes, which could lead to increased adoption across various engineering sectors [11][12] - Synopsys' recent acquisition of Ansys highlights its ambition to lead in multiphysics simulation, which is relevant across multiple industries beyond semiconductors [11][12] - The potential for significant growth in simulation demand is noted, especially if industries shift towards virtual-first workflows due to enhanced computational capabilities [12][25] Group 4: Customer Integration - The integration of accelerated workflows into customer environments remains a key focus, with Synopsys emphasizing its existing relationships across various sectors [14][15] - The specifics of how Synopsys will package and deliver its accelerated tools are still unclear, raising questions about pricing and deployment models [14][15] - NVIDIA's hardware is expected to be well-suited for these workloads, while cloud deployment is seen as a critical avenue for customers lacking high-density computing resources [15][17] Group 5: Neutrality and AI Integration - Concerns about potential bias towards NVIDIA hardware due to the investment were addressed, with both companies affirming that Synopsys' tools will continue to support multiple hardware environments [17][18] - The role of AI in engineering workflows is positioned as a complementary layer rather than a replacement for traditional solvers, emphasizing the need for verified numerical methods [19][20] - AI is expected to enhance design exploration and automate repetitive tasks, but physical solvers will remain foundational in production workflows [20][21]

Core Insights - The article emphasizes that computational power is becoming the "first productive force" in the era of accelerated AI large models, with China's intelligent computing scale expected to grow by 74.1% year-on-year in 2024 [1] - The industry is facing significant challenges, including the "memory wall," "process wall," and "interconnect wall," prompting rapid advancements in technologies such as Chiplet advanced packaging, heterogeneous computing, RISC-V architecture innovation, and distributed clusters [1] - A live forum titled "Building the AI Computing Ecosystem in China" is being organized to address these challenges, featuring key players in the semiconductor industry [1] Group 1: Event Details - The live forum will take place on November 20, 2025, from 14:00 to 16:00 [2] - The event will be accessible via a live streaming platform, with prior registration encouraged [2][8] Group 2: Technical Challenges - In the EDA tools layer, AI-assisted design is crucial for ensuring that domestic AI computing remains autonomous and controllable [6] - The Chiplet architecture layer faces new challenges in system verification, interconnectivity, and standardization across different processes and packages [6] - The computing fusion layer is characterized by a diverse landscape of CPU, GPU, NPU, FPGA, DPU, and emerging architectures like RISC-V, necessitating intelligent collaboration for both Scale-Up and Scale-Out [6] - The ecosystem co-construction layer highlights the need for a closed-loop ecosystem that integrates EDA, Chiplet, NPU, and cloud services, which is still under development [6] Group 3: Roundtable Discussion Topics - The roundtable will discuss how to initiate breakthroughs in autonomous computing systems [7] - It will explore the construction of an evolving computing architecture from Chiplet to system [7] - The discussion will address how to achieve collaborative advancement in a multi-faceted computing ecosystem [7] - It will also focus on igniting collaboration between upstream and downstream players in the industry chain to enhance global competitiveness in AI [7] Group 4: Event Participation - The ICCAD 2025 event is expected to gather over 8,000 industry professionals, 2,000 IC companies, and 300 service providers from the IC industry [7] - For those unable to attend in person, there will be opportunities to virtually explore the event and witness the latest industry trends [7]

Core Viewpoint - The article discusses the significant shifts in the artificial intelligence (AI) industry, highlighting the rapid advancements and the impact on market dynamics, particularly focusing on the performance of various companies in the sector [3]. Group 1: AI Industry Dynamics - The AI revolution is reshaping industry landscapes, with breakthroughs in technology creating substantial market excitement [3]. - Companies like Nvidia have seen their market capitalization soar past $1 trillion due to their dominance in computing chips [3]. - AMD has experienced a remarkable stock price increase, doubling within a year due to its advancements in heterogeneous computing [3]. Group 2: Company Performance - A notable company, one of the earliest entrants in the AI field in China, has seen its market value decline from a peak of 640 billion yuan in early 2021 to just over 300 billion yuan, resulting in a loss of 340 billion yuan [4].

Core Viewpoint - A well-known semiconductor company, Muxi Integrated Circuit (Shanghai) Co., Ltd., is set to undergo an IPO review by the Shanghai Stock Exchange on October 24, 2025 [1] Company Overview - Muxi Integrated Circuit was established in September 2020 in Shanghai and has established wholly-owned subsidiaries and R&D centers in multiple cities including Beijing, Nanjing, Chengdu, Hangzhou, Shenzhen, Wuhan, and Changsha [3] - The core team of the company has an average of nearly 20 years of end-to-end R&D experience in high-performance GPU products [3] - The company focuses on providing full-stack GPU chips and solutions for heterogeneous computing, applicable in advanced fields such as intelligent computing, smart cities, cloud computing, autonomous driving, digital twins, and the metaverse [3] Financial Information - The latest external equity financing corresponds to a post-investment valuation of 21.071 billion yuan [3] - Muxi Integrated Circuit plans to issue no more than 40.1 million A-shares, aiming to raise 3.904 billion yuan [3] - The funds will be allocated as follows: 2.459 billion yuan for the R&D and industrialization of new high-performance general-purpose GPUs, 453 million yuan for the R&D and industrialization of next-generation AI inference GPUs, and 991 million yuan for high-performance GPU technology R&D targeting advanced fields and emerging application scenarios [3] - For the fiscal year 2024, the company reported revenues of 743 million yuan and a loss of 1.409 billion yuan, with the main revenue source being the sales of the Xiyun C500 series training and inference integrated chips [3] Shareholding Structure - As of the signing date of the prospectus, the founder Chen Weiliang controls 22.94% of the voting rights of the company, making him the actual controller [4] - The shareholder list includes prominent investors such as private equity mogul Ge Weidong and his Chaos Investment, as well as firms like Matrix Partners, Helix Capital, and Sequoia Capital [5] - Notably, some listed companies are also "shadow shareholders" of Muxi Integrated Circuit, including Zhongshan Public Utilities, which has made early investments through a renewable energy fund [5]