1月5日，拉斯维加斯。黄仁勋再度向全球抛出"重磅炸弹"——Rubin。其训练性能是Blackwell的3.5倍， AI软件运行性能飙升5倍；但推理成本，仅为前代的1/10。但这种惊人的表现，却是被逼出来的。实际 上，TPU和可重构数据流架构（RPU）的崛起，正凶猛侵蚀英伟达的霸权。 去年11月，Meta拟采用谷歌TPU的传闻一出，英伟达一夜蒸发数千亿美元。焦头烂额的黄仁勋，不得不 光速出手，将可重构数据流芯片公司Groq收入囊中。如今，Groq的并购与Rubin的发布，共同指明了AI 芯片的"收敛时刻"：更高性能的通用芯片，才是所有人要奔赴的终极战场。 疯抢高阶TPU 2025年圣诞节，英伟达豪掷200亿美元（约1400亿人民币），买了一家"非GPU"公司——Groq。英伟达 以其持有现金三分之一的巨资，打包买走了Groq的核心技术、图纸和核心人才。 高阶超越之路 三个月前，Groq的估值还仅为69亿美元。如今，黄仁勋抛出3倍溢价，对Groq志在必得。而Groq主攻 的，是特有的LPU芯片技术，即用软件定义硬件的"可重构数据流架构"。这种独特的设计，能让LPU在 处理大模型时，实现Token"瞬时"+"准时" ...

转自：新华财经 1月5日，黄仁勋再度向全球抛出"重磅炸弹"Rubin，其训练性能是Blackwell的3.5倍，AI软件运行性能飙升5倍；但推理成 本，仅为前代的1/10。但这种惊人的表现是被逼出来的，TPU和可重构数据流架构（RPU）的崛起，正在侵蚀英伟达的统 治地位。 去年11月，Meta拟采用谷歌TPU的传闻一出，英伟达一夜蒸发数千亿美元。焦头烂额的黄仁勋，不得不光速出手，将可重 构数据流芯片公司Groq收入囊中。如今，Groq的并购与Rubin的发布，共同指明了AI芯片的"收敛时刻"：更高性能的通用 芯片，才是所有人要奔赴的终极战场。 疯抢高阶TPU 2025年圣诞节，英伟达豪掷200亿美元（约1400亿人民币），买了一家"非GPU"公司——Groq。英伟达以其持有现金三分 之一的巨资，打包买走了Groq的核心技术、图纸和核心人才。 三个月前，Groq的估值还仅为69亿美元。如今，黄仁勋抛出3倍溢价，对Groq志在必得。而Groq主攻的，是特有的LPU芯 片技术，即用软件定义硬件的"可重构数据流架构"。这种独特的设计，能让LPU在处理大模型时，实现Token"瞬时"+"准 时"的吞吐，超越GPU、TP ...

Core Insights - Qingwei Intelligent, a leading company in reconfigurable AI chips, has completed a C-round financing of over 2 billion RMB, led by Beijing state-owned enterprise Jingneng Group, with participation from nearly twenty other institutions [1][12] - The company is preparing for an IPO, targeting the Sci-Tech Innovation Board, aiming to become the first domestic "non-GPU" new architecture chip stock [1][12] Financing Focus - The funds will be directed towards the iteration of core technologies for the next generation of reconfigurable chips, large-scale deployment in intelligent computing centers, and the recruitment of high-end R&D talent [3][14] Technological Advantages - Qingwei's CGRA (Coarse-Grained Reconfigurable Array) architecture is recognized as a promising fourth type of general-purpose chip, offering significant cost and efficiency advantages over traditional GPU solutions [3][14] - The latest mass-produced chip, TX81, achieves 4 PFLOPS of single-machine computing power, with costs reduced by 50% compared to GPU solutions and energy efficiency improved by three times [3][14] Performance Metrics - The CGRA architecture provides a peak power consumption of less than 1 mW per core, with operational power as low as 0.7 mW per core, significantly outperforming advanced ultra-low-power CPUs [5][16] - The architecture allows for nanosecond-level dynamic reconfiguration, enabling a single chip to function as multiple specialized chips, achieving over 90% hardware utilization [6][16] Market Position - Qingwei Intelligent ranks sixth in domestic AI accelerator card shipments as of the first half of 2025, maintaining a strong position in the first tier of the market [11][23] - The recent financing has the backing of both municipal and district-level state-owned capital, providing synergistic computing resources and support from the capital's digital economy policies [11][23] IPO Prospects - Since its establishment in 2018, Qingwei has completed seven rounds of financing, attracting strategic investors such as Baidu, Ant Group, SK Hynix, and SenseTime, with the latest 2 billion RMB round being crucial for its upcoming IPO [12][24]

Core Insights - Recently, AI chip company Qingwei Intelligent completed over 2 billion RMB in Series C financing, led by Beijing state-owned enterprise Jingneng Group, with participation from various investors [1] - The financing will focus on three areas: next-generation reconfigurable chip development, implementation of intelligent computing scenarios, and high-end talent recruitment [1] - The company has initiated preparations for an IPO, aiming to become the first listed benchmark enterprise in the domestic "non-GPU" new architecture chip sector [1][2] Company Overview - Qingwei Intelligent is a leading company in reconfigurable computing chips, providing chip products and solutions that extend from edge to cloud [2] - The core technology team comes from Tsinghua University Microelectronics Institute, with 13 years of experience in chip R&D, and has received multiple awards including the National Technology Invention Award [2] - The company's chip products began mass production in the first half of 2019, with expected shipments nearing 10 million units [2] Product Development - Qingwei's reconfigurable AI chips have been deployed in over ten large-scale intelligent computing centers across the country, with cumulative orders for computing power cards expected to exceed 20,000 by 2025 [2] - The total shipment of reconfigurable chips is projected to surpass 30 million units, establishing a strong closed loop from technology to market [2] Product Series - The company offers several chip series, including the TX5 series aimed at edge intelligent scenarios, which features flexibility, low power consumption, and high efficiency [4] - The TX8 series targets the cloud market, utilizing a reconfigurable data flow architecture that allows direct interconnection between chips and systems, effectively solving power expansion issues and reducing costs [7] - The latest TX81 module achieves a computing power of 512 TFLOPS (FP16), with the REX1032 training and inference server reaching 4 PFLOPS, supporting deployment of large models with over a trillion parameters [7] Strategic Initiatives - This year, Qingwei aims to leverage its product and market advantages to build a domestic autonomous AI ecosystem, collaborating with the Beijing Academy of Artificial Intelligence to establish a joint laboratory [8] - The company is also actively participating in the open-source community and has become a leading unit in the domestic chip software ecosystem [8] - Qingwei has partnered with several AI innovation institutions to establish the "Beijing Reconfigurable Computing Hardware and Software Collaborative Technology Innovation Center," focusing on deep integration of industry, academia, and research [10]

Core Insights - The 2025 Integrated Circuit Development Forum (ICCAD-Expo) held in Chengdu highlighted the optimistic sales forecast for China's chip design industry, projecting sales to exceed $100 billion for the first time by 2025, indicating a return to high growth [1][3] - The event gathered over 2,000 renowned companies in the integrated circuit industry, including major players like Samsung, Siemens, and TSMC, showcasing the latest technologies [1][3] - The forum emphasized the significant growth potential of non-GPU AI chips, with projections indicating that the market share of reconfigurable chips in non-GPU server markets will reach 30% by mid-2025 and grow to 50% by 2028 [3][5] Industry Developments - The theme of this year's ICCAD-Expo was "Open Innovation, Achieving the Future," attracting over 6,300 industry experts to discuss the development trends of the integrated circuit industry [3] - The Chinese semiconductor industry is expected to experience a new wave of growth, potentially reaching a scale of 1 trillion yuan by 2030, driven by advancements in AI and other new technologies [3][5] - Clear Micro Intelligence, a domestic reconfigurable computing chip company, presented insights on the strategic development opportunities in the non-GPU chip architecture during the forum [5][6] Company Highlights - Clear Micro Intelligence's CEO, Wang Bo, was awarded the "2025 IC Design Annual Entrepreneur" title for his strategic vision and commitment to promoting the localization of new AI chip architectures [6][10] - The company has successfully mass-produced the TX81 chip, which is designed for large-scale AI applications, achieving a 50% cost reduction compared to competitors and a threefold increase in energy efficiency [6][7] - Since the mass production of its products, Clear Micro Intelligence has established large-scale intelligent computing clusters across multiple regions in China, with over 20,000 orders for reconfigurable computing cards [7]

Group 1 - The global AI chip market is witnessing a shift towards data flow architecture, with companies like SambaNova and Groq achieving significant valuations of $5 billion and $6 billion respectively [1] - Clear Microelectronics, originating from Tsinghua University, has successfully developed and mass-produced data flow reconfigurable chip technology, positioning itself as a leader in this emerging field [1][2] - The founder of Clear Microelectronics, Wang Bo, emphasizes the need for innovation beyond traditional GPU architectures to overcome limitations in technology and materials, advocating for a "leapfrog" approach similar to the automotive industry's transition to electric vehicles [2] Group 2 - Clear Microelectronics' first "new computing power" chip, TX81, has achieved over 20,000 orders and established intelligent computing centers across multiple regions in China within just six months of its launch [2] - Investment institutions are increasingly recognizing the value of new computing power, with significant investments from major funds indicating a strong market trend towards data flow architecture [3] - The transition to data flow architecture is seen as a critical signal for achieving self-sufficiency in the computing power industry, with support from initiatives like ChatGPT and DeepSeek3.1 [3]

Core Viewpoint - The article discusses the emergence of reconfigurable chips as a potential disruptor in the AI chip market, contrasting them with traditional GPU architectures dominated by companies like NVIDIA. It highlights the innovative approach taken by Qingwei Intelligent, led by founder Wang Bo, in developing reconfigurable chips that can dynamically allocate computing resources, thus offering a competitive edge in performance and cost efficiency [4][5][6]. Group 1: Reconfigurable Chip Technology - Reconfigurable chips represent a fundamentally different computing paradigm compared to traditional GPUs, characterized by instruction-driven shared storage versus instructionless configuration and data flow-driven processing [7][9]. - Wang Bo's analogy compares GPUs to a straight railway track, while reconfigurable chips allow for multiple "switches" to adapt to various tasks dynamically [6][8]. - The technology aims to address the limitations of existing architectures, particularly in the context of AI's growing computational demands, as traditional architectures struggle to keep pace with the required performance [10][11]. Group 2: Market Position and Strategy - Qingwei Intelligent's first commercial product, the TX81 chip, achieved significant market traction, with over 20,000 orders and deployment in various intelligent computing centers within six months of launch [17][18]. - Wang Bo emphasizes the need for a "5x cost-performance advantage" to compete effectively against established players like NVIDIA, which includes superior performance and lower operational costs [18][29]. - The company plans to leverage 3D storage technology in its next-generation TX8 series chips to enhance performance and achieve the targeted cost-performance ratio [19][30]. Group 3: Competitive Landscape - The reconfigurable chip technology is not unique to Qingwei Intelligent, as other companies like Google with its TPU and emerging startups are also exploring similar architectures, indicating a growing trend towards alternative chip designs outside the traditional GPU ecosystem [20][34]. - Wang Bo acknowledges the competitive challenge posed by established companies and the necessity for innovative approaches to carve out market share [26][28]. - The article highlights the importance of building an ecosystem around the new architecture, including compatibility with existing frameworks like CUDA and open-source initiatives, to facilitate user adoption [32][33].

Core Viewpoint - The article discusses the emergence of reconfigurable chips as a potential disruptor to the dominance of traditional GPU architectures, particularly those developed by Nvidia and Intel. Wang Bo, the founder of Qingwei Intelligent, represents a new approach that diverges from established paths in the AI chip market [1][2]. Group 1: Reconfigurable Chip Technology - Reconfigurable chips allow for dynamic configuration of computing resources, contrasting with GPUs that follow a fixed instruction-driven model. This flexibility enables multiple tasks to be performed without the need for extensive data transfer protocols [2][3]. - The architecture of reconfigurable chips moves away from the traditional von Neumann architecture, focusing on data flow-driven computation rather than instruction-driven processes [5][6]. Group 2: Market Position and Strategy - Wang Bo emphasizes that to compete with established players like Nvidia, companies must not follow the same paths, as this leads to inevitable failure. Instead, innovation and differentiation are crucial [2][24]. - Qingwei Intelligent aims to achieve a "5x cost-performance advantage" over competitors, which includes superior performance and lower costs, to attract customers and encourage them to switch from existing models [14][24]. Group 3: Product Development and Challenges - The company faced significant challenges in transitioning from laboratory technology to commercial products, requiring a focus on reliability, compatibility, and customer needs [8][21]. - After initial setbacks in consumer electronics, Qingwei Intelligent pivoted to focus on AI-centric chip development, leading to the successful launch of the TX81 chip, which has already garnered over 20,000 orders [11][14]. Group 4: Future Prospects and Innovations - The next generation of TX8 series chips will incorporate "3D storage" technology to enhance performance and achieve the promised cost-performance advantage [15][25]. - The company is also working on compatibility with existing ecosystems, including Nvidia's CUDA and open-source frameworks like Triton, to facilitate easier adoption of their technology by potential customers [28][29].

Core Viewpoint - The emergence of the TX81 chip represents a significant breakthrough in China's technology landscape, potentially rewriting the narrative of domestic chip development and AI capabilities [3][6][9]. Group 1: Technological Innovation - The TX81 chip is a dynamic reconfigurable chip (RPU) that allows efficient computation without complex instruction set exchanges, adapting to various tasks such as voice, image, and large models [7][8]. - This chip embodies four core advantages: high energy efficiency, high concurrency, high scalability, and high cost-effectiveness [8]. - The architecture innovation of the TX81 chip is seen as a necessary path for China to overcome the limitations of traditional chip development and compete in the global AI landscape [5][9]. Group 2: Market Context - The traditional growth curve of chips is becoming inadequate in the face of explosive AI computing demands, necessitating a shift towards innovative architectures [5]. - The success of the TX81 chip is likened to the transformation seen in the Chinese automotive industry, where companies like BYD and Xiaomi have achieved breakthroughs through electrification and smart technology [10][13]. Group 3: Company Background - The founding of Qingwei Intelligent was driven by a vision to commercialize reconfigurable chips after the founder faced challenges in selecting suitable chips for machine vision applications [16][22]. - The company started with a small team in a modest office, emphasizing a culture of innovation and resilience despite initial hardships [29][32]. Group 4: Development Challenges - The development of the TX81 chip involved significant challenges, including achieving a large chip area of 800 square millimeters and enhancing computing efficiency through reconfigurable technology [35]. - The team faced numerous technical hurdles, including the need for custom clock devices and programming for multi-chip interconnections, which required original solutions [36][40]. Group 5: Future Prospects - Qingwei Intelligent aims to continue its momentum with the upcoming TX82 chip, targeting mass production by 2026, as part of its commitment to advancing China's AI capabilities [51][53]. - The collaboration with other domestic firms, such as Zhiyu, highlights a growing ecosystem focused on original technology development in China [53].