当AI时代算力集群规模正逐步从万卡向十万卡、百万卡甚至千万卡升级时，一支中国团队悄然另辟蹊径。 今年10月，北京大学人工智能研究院/集成电路学院双聘助理教授孙仲与北京大学集成电路学院蔡一茂教授、王宗巍助理教授率领的团队成功研制出基于 阻变存储器的高精度、可扩展模拟矩阵计算芯片，在全球范围内首次将模拟计算的精度提升至24位定点精度，让未来同等任务下使用更少的计算卡成为可 能。 这是一种完全不同于目前所有商用量产芯片的新型芯片，计算精度从1%跃升至千万分之一；可以支撑6G、具身智能及AI大模型训练等多个前沿场景；更 重要的是，它可在28纳米及以上成熟工艺量产，绕开光刻机"卡脖子"环节。 事实上，AI大模型、具身智能、6G等应用背后都是矩阵计算，AI推理是做矩阵乘法，AI训练是在解矩阵方程。而英伟达的崛起正是得益于GPU（图形处 理器）很擅长做矩阵计算。 但如今摩尔定律趋于终结，晶体管很难再微缩，所以业界现在只能横向堆计算卡：少则百卡，多则万卡、十万卡。但这样的方式我认为是不可持续的—— 能耗、碳排放均呈指数级上升，与国家"双碳"目标相悖。因此，我认为需要探索一种不同的计算范式，即模拟（类比）计算。 模拟计算并非全 ...

Core Insights - A Chinese research team has developed a new type of chip based on resistive random-access memory (RRAM) that achieves a precision of 24-bit fixed-point accuracy in analog matrix computations, marking a significant advancement in computational efficiency and energy consumption for AI applications [2][12][15] - This chip can support various cutting-edge applications, including 6G communication, embodied intelligence, and AI model training, while being produced using mature 28nm technology, thus avoiding reliance on advanced lithography processes [2][4][10] Technology Overview - The new chip represents a departure from traditional digital computing paradigms, which rely on binary logic and silicon-based transistors, to a more efficient analog computing approach that directly utilizes physical laws for calculations [4][6][15] - The precision of analog computing has been significantly improved, reducing relative error from 1% to one part in ten million (10⁻⁷), which is crucial for large-scale computations where errors can accumulate exponentially [8][12][15] Innovation Highlights - The chip's innovations include the use of RRAM as a core component, a novel feedback circuit design that minimizes energy consumption while enhancing accuracy, and the implementation of classic iterative optimization algorithms for efficient matrix equation solving [15][16] - The chip's architecture allows for high-speed, low-power solutions to matrix equations, making it suitable for applications that require rapid computations, such as second-order training methods in AI [19][21] Application Potential - The chip is particularly well-suited for medium-scale applications, such as AI model training and 6G MIMO systems, where it can outperform traditional digital chips [18][25] - Future plans include scaling the chip's matrix size from 16x16 to 128x128 within two years, with aspirations to reach 512x512, which would enhance its applicability in various computational scenarios [25][26] Strategic Value - This development provides China with a potential alternative to reliance on advanced processes and NVIDIA GPUs, positioning the country favorably in the global computational landscape [10][11] - The successful demonstration of this new computing paradigm is seen as a critical step towards addressing future computational demands, emphasizing the need for ongoing investment in technology and infrastructure [11][26]

Group 1: DeepSeek AI Model - The Chinese AI company DeepSeek launched its open-source model DeepSeek-R1, which has gained global attention due to its low training costs and high performance in tasks like mathematical reasoning and code generation [2][3] - DeepSeek-R1's core competitiveness lies in its systematic innovation in computational efficiency, demonstrating that top-tier reasoning capabilities can be achieved without massive labeled data, significantly reducing training costs [2][3] - The model's open-source approach breaks technological monopolies, allowing developers worldwide to participate in its ecosystem, which has attracted hundreds of thousands of developers [2] Group 2: Nuclear Fusion and Quantum Computing - China's "artificial sun," the EAST device, achieved a world record by maintaining a plasma temperature of 100 million degrees Celsius for 1000 seconds, marking a significant step towards practical nuclear fusion energy [4] - The superconducting quantum computing prototype "Zuchongzhi 3" was developed, showcasing a computational speed that is a trillion times faster than the current fastest supercomputers, indicating a major advancement in quantum computing capabilities [5][6] Group 3: Advanced Materials and Brain-Computer Interfaces - A research team successfully created large-area two-dimensional metallic materials, marking a significant breakthrough in the field of two-dimensional materials [7] - China initiated its first invasive brain-computer interface clinical trial, allowing participants to control devices through thought, utilizing advanced flexible neural electrodes that minimize brain tissue damage [8][9] Group 4: Lunar Exploration and Plant Biology - The Chang'e 6 mission revealed the evolutionary history of the moon's far side, providing insights into volcanic activity and magnetic fields, which are crucial for understanding lunar geology [11][12] - A research team uncovered the molecular mechanisms behind how a single plant cell can develop into a complete plant, addressing a long-standing scientific question in plant biology [13] Group 5: Technological Innovations in Computing - Researchers developed a high-precision, scalable analog matrix computing chip, achieving digital-level precision in analog computing, which could revolutionize computational tasks in AI and communications [14][15] Group 6: National Strategic Initiatives - The 20th Central Committee of the Communist Party of China emphasized the importance of technological innovation in its strategic planning for the next five years, aiming to enhance China's technological self-reliance and drive new productive forces [16][17] Group 7: Military Advancements - China's first electromagnetic catapult aircraft carrier, Fujian, was commissioned, representing a leap in naval technology by utilizing advanced electromagnetic launch systems, enhancing operational capabilities [18]

Group 1: Artificial Intelligence Developments - The Chinese AI company DeepSeek launched the open-source model DeepSeek-R1, which has gained global attention due to its low training costs and high performance in tasks like mathematical reasoning and code generation [2] - DeepSeek-R1's core competitiveness lies in its systematic innovation in computational efficiency, achieving top-tier reasoning capabilities without the need for massive labeled data [2][3] - The model's open-source approach aims to break technological monopolies, allowing developers worldwide to participate in its ecosystem [2] Group 2: Nuclear Fusion Research - China's "artificial sun," the EAST device, achieved a world record by maintaining a plasma temperature of 100 million degrees Celsius for 1000 seconds, marking a significant step towards practical nuclear fusion [4] - The high-confinement operation mode is crucial for future fusion reactors, indicating that the experiment has successfully simulated the necessary conditions for sustained fusion [4] Group 3: Quantum Computing Advancements - The "Zu Chongzhi No. 3" superconducting quantum computing prototype was developed, demonstrating a computational speed that surpasses the fastest supercomputers by trillions of times [5][6] - This prototype achieved the highest level of quantum computing superiority, showcasing its potential for various applications in quantum error correction and simulation [6] Group 4: Material Science Innovations - A research team successfully created large-area two-dimensional metallic materials, marking a significant advancement in the field of material science [7] - This breakthrough allows for the production of ultra-thin metals, potentially opening new avenues for research in two-dimensional materials [7] Group 5: Brain-Computer Interface Trials - China initiated its first invasive brain-computer interface clinical trial, positioning itself as the second country globally to enter this phase of technology [8][9] - The trial utilizes flexible neural electrodes that minimize damage to brain tissue, enhancing the safety and effectiveness of the procedure [8] Group 6: Lunar Exploration Findings - The Chang'e 6 mission revealed the evolutionary history of the moon's far side, providing insights into volcanic activity and the moon's magnetic field [11][12] - This research fills a significant gap in lunar studies, highlighting the moon's geological history and the impact of large-scale collisions on its evolution [11][12] Group 7: Agricultural Biotechnology Breakthroughs - A research team unveiled the molecular mechanisms behind how a single plant cell can develop into a complete plant, addressing a long-standing scientific challenge [13] - This discovery could pave the way for advancements in agricultural biotechnology, particularly in overcoming regeneration bottlenecks [13] Group 8: Computing Architecture Innovations - Researchers developed a high-precision, scalable analog matrix computing chip, achieving precision comparable to digital computing systems [14][15] - This innovation addresses the challenges of computational efficiency in AI and 6G communications, marking a significant breakthrough in computing paradigms [15] Group 9: National Strategic Initiatives - The 20th Central Committee of the Communist Party of China emphasized the role of technological innovation in driving economic development in its 14th Five-Year Plan [16][17] - The plan outlines specific strategies to enhance original innovation and integrate technological advancements with industrial development [16][17] Group 10: Military Advancements - China's first electromagnetic catapult aircraft carrier, the Fujian, was commissioned, representing a leap in naval technology with its advanced launch capabilities [18] - This carrier enhances operational capabilities and signifies a transition to a new era in the Chinese navy, showcasing advancements in military technology [18]

作者丨冯汝梅 编辑丨关雎 人工智能赛道又一个惊人融资事件诞生。 2025年12月8日，由前Databricks人工智能主管Naveen Rao创立仅两个月的AI神经形态计算芯片初创公司——Unconventional AI，在种子轮融资中筹集了 4.75亿美元（约合33亿元），投后估值飙升至45亿美元（约合318亿元）。这一融资规模不仅远超常规种子轮水平，甚至超过多数初创企业的C轮融资，创 下AI芯片领域早期融资纪录之一。 此次融资由Andreessen Horowitz（a16z）和Lightspeed Venture Partners共同领投，其他投资者包括Lux Capital和DCVC。Rao的前雇主Databricks以及亚马逊 创始人杰夫·贝索斯（Jeff Bezos）也参与了本轮投资。Rao本人也按与其他投资者相同的条款投资了1000万美元。 值得注意的是，这笔融资还只是Unconventional AI更大规模融资计划的第一步。根据Rao透露，公司计划后续融资总额可能高达10亿美元。 作为一家尚无产品、成立仅数周的初创公司，Unconventional AI的核心竞争力首先源于其豪华的创始团队 ...

Core Insights - Unconventional AI, a startup founded by Naveen Rao, raised $475 million in seed funding, achieving a post-money valuation of $4.5 billion, marking a record in early-stage financing within the AI hardware sector [3][4]. - The company aims to develop next-generation digital computing by designing simulation chips inspired by neuroscience principles, addressing the energy consumption challenges faced by traditional AI computing [15][19]. Company Overview - Unconventional AI was established just two months prior to its significant funding round, with a founding team that includes experts from MIT, Stanford, and former Google engineers, providing a comprehensive capability chain from theory to application [5][7]. - Rao's previous entrepreneurial successes include Nervana Systems, which was acquired by Intel for approximately $400 million, and MosaicML, which was sold to Databricks for $1.3 billion [12][14]. Technological Vision - The company seeks to redefine AI computing hardware architecture by creating high-efficiency simulation chips tailored for AI workloads, diverging from the traditional reliance on GPUs [17][20]. - Unconventional AI's approach contrasts with the prevailing "scaling laws" in AI development, which emphasize increasing computational power and data size, by focusing on energy efficiency and the probabilistic nature of AI tasks [18][24]. Industry Context - The rise of "Neo-Lab" startups, like Unconventional AI, reflects a shift in the AI landscape where founders with proven track records are attracting significant investment for long-term foundational research rather than immediate product commercialization [25][26]. - The funding environment is increasingly favoring companies that challenge existing paradigms in AI development, as evidenced by the substantial valuations of similar startups [28].

Core Insights - A research team from Peking University has developed a high-precision, scalable analog matrix computing chip based on resistive random-access memory (ReRAM), achieving analog computing precision comparable to digital systems [2][4] - The chip significantly enhances computational throughput and energy efficiency, reportedly improving performance by 100 to 1000 times compared to current top digital processors (GPUs) when solving large-scale MIMO signal detection problems [2][4] - This technological breakthrough addresses global challenges of slowing digital computing power growth and rising energy consumption, offering a new solution for critical fields such as AI and autonomous driving [2][6] Analog vs. Digital Computing - Analog computing was once the dominant form of computation but was replaced by digital computing due to precision and scalability limitations [4] - The new chip aims to resolve the precision issues of analog computing, achieving a relative error as low as 10^-7 after 10 iterations for a 16x16 matrix inversion, which meets the needs of most scientific calculations and AI training [4][9] - The chip's performance surpasses high-end GPU single-core performance when solving 32x32 matrix inversion problems, and achieves over 1000 times the throughput of top digital processors for 128x128 matrices [4][7] Advantages of the New Chip - The chip utilizes a "compute-storage integration" approach, eliminating the need for data to be converted into binary streams, thus reducing energy consumption associated with data transfer [5][6] - The low power consumption and high energy efficiency of the analog computing chip align well with the energy management needs of electric vehicles, potentially enhancing their driving range [7][9] - The chip is expected to significantly reduce the training time for AI models, particularly in autonomous driving, where traditional GPUs may take hours to complete tasks that the new chip could finish in minutes [7][10] Industry Perspectives - While the research results are promising, industry experts express caution regarding the practical application of the technology, particularly in the automotive sector, where reliability and durability under harsh conditions are critical [9][10] - The transition from laboratory to industrial application faces challenges such as cost, supply chain maturity, and the need for robust manufacturing processes for the new chip technology [9][10] - The current state of resistive memory technology is still in the experimental phase, with material consistency and reliability needing further development to meet automotive standards [10]

Group 1: Technology Innovations - Peking University has developed a new type of analog computing chip that significantly improves computing efficiency and reduces energy consumption, enhancing applications in artificial intelligence [2] - ByteDance's Seed team launched the 3D generative model Seed 3D1.0, capable of generating high-quality simulation-level 3D models from a single image, addressing current limitations in physical interaction and content diversity [4] Group 2: Market Trends and Insights - Multiple insurance investment institutions recommend prioritizing technology sectors, especially in Hong Kong stocks, as the trading congestion has eased and technology stocks are seen as more attractive in terms of price-to-earnings growth ratios [11] - QFII has increased stakes in 29 companies, focusing on advanced manufacturing sectors with strong performance, indicating a positive outlook for these industries [12] Group 3: Industry Developments - Alibaba's first self-developed AI glasses, Quark AI glasses, are set to launch with a pre-sale price of 3,999 yuan, featuring advanced functionalities [5] - The AI comic industry is experiencing rapid growth, with a significant increase in production and a projected market scale exceeding 20 billion yuan by 2025 [7]

Core Insights - Chinese scientists have achieved a significant breakthrough in computing architecture with the development of a high-precision, scalable analog matrix computing chip based on resistive random-access memory (ReRAM) [1] - This new technology enhances the precision of analog computing to 24-bit fixed-point accuracy, with computational throughput and energy efficiency surpassing current top GPUs by a factor of 100 to 1000 [1] - The technology aims to liberate computing power by integrating data computation and storage, potentially disrupting the long-standing dominance of digital computing [1] Industry Impact - The mainstream CPUs and GPUs currently in the market utilize digital chips and the von Neumann architecture, which separates computing and storage functions [1] - The analog computing advantage lies in eliminating the need to convert data into binary streams and avoiding process data storage, thus merging computation and storage [1] - Industry experts believe this technology could pave the way for a new era of ubiquitous and energy-efficient computing power [1] Related Companies - Relevant A-share concept stocks include Shengbang Co., Ltd. and Changdian Technology [2]

Core Viewpoint - The article discusses the successful development of a high-precision, scalable analog matrix computing chip based on resistive random-access memory (RRAM) by a research team from Peking University, which achieves computational efficiency and energy performance significantly superior to current top digital processors, with improvements ranging from 100 to 1000 times [1][9]. Group 1: Analog Computing Concept - Analog computing allows for direct representation of mathematical values using continuous physical quantities, such as voltage, eliminating the need for binary conversion [4][5]. - Historically, analog computers were widely used before being replaced by digital computers due to precision limitations, which this new research aims to address [5][7]. Group 2: Technical Advantages - The new chip integrates data computation and storage, removing the need for binary data conversion and enabling a more efficient processing method [7]. - The research focuses on solving matrix equations, particularly matrix inversion, which is crucial for AI training, and demonstrates significant performance improvements over traditional GPUs [7][9]. Group 3: Performance Metrics - The team achieved a precision of 24-bit for 16x16 matrix inversion, with relative errors as low as 10⁻⁷ after 10 iterations [9]. - For larger matrices, the chip's performance exceeds that of high-end GPUs, achieving over 1000 times the throughput of top digital processors for 128x128 matrix problems, completing tasks in minutes that would take traditional GPUs a day [9]. Group 4: Future Applications - The chip is expected to serve as a powerful complement in the AI field, particularly in computational intelligence applications such as robotics and AI model training [11]. - The future landscape will likely see coexistence between CPUs, GPUs, and this new analog computing chip, enhancing the overall computational efficiency in energy-intensive tasks [11].