证券研究报告：计算机|点评报告 发布时间：2025-08-27 行业投资评级 强于大市|维持 行业基本情况 | 收盘点位 | | 5804.0 | | --- | --- | --- | | 52 | 周最高 | 5804.0 | | 52 | 周最低 | 2805.53 | 行业相对指数表现（相对值） -5% 6% 17% 28% 39% 50% 61% 72% 83% 94% 2024-08 2024-11 2025-01 2025-04 2025-06 2025-08 计算机 沪深300 资料来源：聚源，中邮证券研究所 研究所 分析师:陈涵泊 SAC 登记编号:S1340525080001 Email:chenhanbo@cnpsec.com 分析师:李佩京 SAC 登记编号:S1340525080003 Email:lipeijing@cnpsec.com 分析师:王思 SAC 登记编号:S1340525080002 Email:wangsi1@cnpsec.com 近期研究报告 《DeepSeek V3.1 提振国产算力和应 用》 - 2025.08.25 中国移动集采大单公布，推理型人工智能部署加 ...

华鑫证券近日发布电力设备行业周报：产业升级上，"国产芯片+液冷技术"双轮驱动重 构算力基建：华为昇腾910C实现规模化部署打破海外垄断，与液冷形成"芯片-散热"方案破 解散热瓶颈，Deepseek-V3.1模型适配新一代国产芯片，加速"芯片-模型"生态闭环；AIDC机 柜功率达20-100kW，液冷因高效低耗成刚性需求，渗透率加速提升。 以下为研究报告摘要： 投资要点 多因素催化，IDC与算力租赁赛道迎来爆发契机 国产算力迎多重利好，投资价值凸显。产业升级上，"国产芯片+液冷技术"双轮驱动重 构算力基建：华为昇腾910C实现规模化部署打破海外垄断，与液冷形成"芯片-散热"方案破 解散热瓶颈，Deepseek-V3.1模型适配新一代国产芯片，加速"芯片-模型"生态闭环；AIDC机 柜功率达20-100kW，液冷因高效低耗成刚性需求，渗透率加速提升。政策驱动下，国有数 据中心国产芯片采购要求超50%，叠加火山引擎60亿元内蒙项目释放产业链需求；英伟达暂 停H20生产刺激国产替代，8月22日中国算力大会或进一步明确政策支持，推动赛道国产化 转型。出海逻辑强化，柴发厂商借海外AI算力基建需求，凭产品优势与成本力出海， ...

克雷西 一水 发自 凹非寺 量子位 | 公众号 QbitAI DeepSeek V3.1发布后，一则官方留言让整个AI圈都轰动了： 新的架构、下一代国产芯片，总共短短不到20个字，却蕴含了巨大信息量。 国产芯片企业股价也跟风上涨，比如寒武纪今日早盘盘中大涨近14%，总市值跃居科创板头名。 半导体ETF，同样也是在半天的时间里大涨5.89%。 （不知道作为放出消息的DeepSeek背后公司幻方量化，有没有趁机炒一波【手动狗 头】） | Cambricon | 其武红 | + | | | | | | | | | | | 每日Ali "股票i | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | | | | | SH 688256 ■ Level1基础行情 ■ 上海交易所 ■ 沪港通标的股票 ■ 科创板 ■ 融资融券标的 | | | | | | | | | | 所属行业 × 半导体 +2.68% > | | | | | | | | | | | | | | | 1164.45元 +128. ...

Core Viewpoint - The article discusses the architecture and data flow of Huawei's Cloud Matrix 384, emphasizing the integration of optical and electrical interconnections in its network design [2][3][9]. Group 1: Data Transmission Layers - The Cloud Matrix 384 includes three main data transmission layers: UB Plane, RDMA Plane, and VPC Plane, each serving distinct roles in data processing and communication [5][7]. - The UB Plane connects all NPU and CPU with a non-blocking full-mesh topology, providing a unidirectional bandwidth of 392GB/s per Ascend 910C [7]. - The RDMA Plane facilitates horizontal scaling communication between supernodes using RoCE protocol, primarily connecting NPUs for high-speed KV Cache transfer [7]. - The VPC Plane connects supernodes to broader data center networks, managing tasks such as storage access and external service communication [7]. Group 2: Optical and Electrical Interconnections - Although the Cloud Matrix 384 is often referred to as a purely optical interconnection system, it also utilizes electrical interconnections for short distances to reduce costs and power consumption [9]. - The article highlights the necessity of both optical and electrical connections in achieving efficient data flow within the system [9]. Group 3: Scale-Up and Scale-Out Calculations - For Scale-Up, each server's UB Switch chip corresponds to a bandwidth of 448GBps, requiring 56 400G optical modules or 28 800G dual-channel optical modules per server [12]. - The ratio of NPUs to 400G optical modules in Scale-Up is 1:14, and to 800G modules is 1:7 [12]. - For Scale-Out, a Cloud Matrix node consists of 12 Compute cabinets, and the optical module demand ratio is approximately 1:4 for NPUs to 400G optical modules [14].

Investment Rating - The report rates the industry as "Outperforming the Market" [8]. Core Insights - The report identifies Huawei's future growth direction over the next decade as focusing on enterprise business, empowering various industries [3]. - Following multiple rounds of sanctions, Huawei has largely achieved a self-controlled full industrial chain [3]. - Huawei's enterprise business is expected to be the most certain growth direction in the coming years, as it aligns with the digital transformation of Chinese enterprises [5][6]. Summary by Sections Huawei's Next Decade of Growth - Huawei is accelerating the construction of a self-controlled soft and hard integrated digital ecosystem, focusing on the digital transformation of various industries [21]. - The company has shifted its revenue focus from overseas markets to domestic demand, with significant growth in the Chinese market since 2008 [22]. Enterprise Business as Core Increment - The enterprise business is a combination of "digital foundation + intelligent upgrade + industry customization," helping various organizations with their digital transformation [35]. - Huawei's enterprise business has expanded from hundreds of billions to a target of 260 billion by 2025, indicating strong growth potential [37]. Industry Corps and Full-Process Empowerment - Huawei has established 21 industry corps to implement its "All Intelligence" strategy, focusing on various sectors such as government, large enterprises, and manufacturing [40][43]. - The report outlines nine core directions for digital transformation across industries, including digital government, large enterprises, oil and gas mining, and education [45]. Market Space for Digital Transformation - The report estimates a massive market space for digital transformation, suggesting that if industrial enterprises allocate just 1% of their revenue to digitalization, it would represent a market space of 1.38 trillion [6][37].

Core Viewpoint - The article discusses the intense competition for AI talent, particularly among Chinese scientists and engineers, highlighting their significant role in the AI industry and the ongoing "talent war" in Silicon Valley [3][5][15]. Group 1: Talent Acquisition - In 2023, after Elon Musk's split from OpenAI, he founded xAI, which prominently features Chinese scientists in its leadership, with a notable 5 out of 12 founding members being of Chinese descent [7][11][12]. - Meta's CEO Mark Zuckerberg is aggressively recruiting Chinese engineers, offering lucrative contracts, including a reported $200 million deal for a key Apple team leader [14][16]. - The competition for top AI talent has led to significant financial incentives, with Meta reportedly offering up to $300 million in total compensation for top researchers [14][16]. Group 2: Chinese AI Development - Chinese researchers are increasingly recognized as core contributors to AI advancements, with a report indicating that 65 out of the top 100 AI experts globally are from China [18]. - From 2010 to 2023, the number of AI patents surged from 3,833 to 122,511, with China holding 69.7% of all authorized AI patents as of 2023 [19]. - The rise of Chinese AI capabilities is attributed to systematic educational and policy support, with over 4500 AI companies currently operating in China [21]. Group 3: Entrepreneurial Ventures - Chinese entrepreneurs are making significant strides in the AI startup scene, with notable acquisitions and innovations, such as Scott Wu's Cognition and the acquisition of Scale AI by Meta [25][26]. - The emergence of companies like Surge AI, led by Edwin Chen, showcases the competitive landscape where Chinese founders are driving innovation without external funding [27]. - The article emphasizes a shift where Chinese scientists and entrepreneurs are not just being recruited but are also becoming key players in shaping the future of AI [34]. Group 4: Global Impact and Future Trends - The article highlights a cultural shift towards a more optimistic view of AI in China, with 83% of Chinese respondents believing AI is beneficial, compared to only 39% in the U.S. [22]. - The return of Chinese scientists to domestic startups is creating a "returning force" that enhances China's AI capabilities in algorithm development and model innovation [34]. - The competitive landscape is evolving, with Chinese AI models like DeepSeek challenging established players like ChatGPT, indicating a significant shift in the global AI market dynamics [28][30].

Core Viewpoint - The article discusses the significance of the H20 graphics card from NVIDIA, particularly its implications for AI model training in China, highlighting the competitive landscape of computing power in the AI era [6][12]. Group 1: H20 Graphics Card Overview - The H20 graphics card, set to be released in late 2023, is designed specifically for AI model inference training and has a performance level of 10% to 15% compared to the previous flagship H100 model [6]. - Despite its high price of 110,000 yuan, the H20 is currently the best-performing graphics card available in China, leading major tech companies to purchase significant quantities [7]. Group 2: Market Dynamics and Sales Restrictions - The H20 was initially banned from sale in China on April 9, causing concern for NVIDIA as it impacted their sales strategy [8]. - To mitigate losses, NVIDIA's CEO sought to lift the sales ban, citing that the H20's performance is no longer considered high-end compared to newer models like the GB200, which is estimated to be over 50 times more powerful [9]. Group 3: Competitive Landscape - The potential for domestic competitors, such as Huawei's Ascend 910C, to fill the market gap if H20 supply is not restored is a significant concern for NVIDIA [10]. - The resumption of H20 sales underscores the ongoing competition for computing power in the AI sector, where the quantity of graphics cards directly correlates with the ability to train superior AI models [12][13].

Core Viewpoint - The competition for computing power, particularly in AI model training, is intensifying, with the availability of graphics cards being a crucial factor in determining success in this arena [1][12]. Group 1: H20 Graphics Card Overview - The H20 graphics card, launched by NVIDIA at the end of 2023, is designed for AI model inference training and has a performance level of 10% to 15% compared to the previous flagship H100 [3]. - Despite the high price of 110,000 yuan (approximately 15.5 million USD), the H20 is currently the best-performing graphics card available in China, leading many tech companies to purchase significant quantities [4]. Group 2: Market Demand and Sales Dynamics - Major companies like ByteDance, Alibaba, and Tencent ordered at least 16 billion USD worth of H20 graphics cards in the first quarter of last year, with ByteDance reportedly stockpiling 100,000 GPU modules [5]. - The sales of H20 were initially banned in China on April 9, causing significant concern for NVIDIA, which prompted efforts to lift the ban [7]. Group 3: Competitive Landscape and Strategic Implications - NVIDIA's CEO is motivated to restore sales of the H20 due to the relatively low performance of the card compared to the latest GB200, which is estimated to be over 50 times more powerful [8]. - If NVIDIA fails to supply the H20, domestic competitors like Huawei's Ascend 910C, which is approaching the performance parameters of the H100, may fill the market gap [9]. - The resumption of H20 sales indicates that the race for computing power in the AI era is far from over, emphasizing the importance of graphics card availability in training superior AI models [10][13].

Core Insights - The report from Minsheng Securities outlines the mid-term investment strategy for the electronics industry, emphasizing the growth of AI-driven applications and the increasing demand for computing power, particularly in the context of Nvidia's strong performance and the rise of self-developed ASICs by CSPs [1][2]. Group 1: Overseas Computing Power - Recent highs in US AI hardware and software stocks are driven by Nvidia's better-than-expected earnings, with long-term growth fueled by AI applications enhancing internet capabilities and increasing inference demand [2]. - The demand for computing power is expected to grow significantly, leading to accelerated product iterations from Nvidia and faster growth for CSP's self-developed ASICs [2]. - Upgrades in computing power rely on two main routes: speed improvements through PCB upgrades and the evolution from traditional optical modules to CPO, and power enhancements via HVDC and supercapacitors [2]. Group 2: Domestic Computing Power - Domestic models like Doubao and DeepSeek are advancing in multi-modal and lightweight capabilities, accelerating the development of domestic large models [3]. - Domestic cloud computing firms are increasing investments in computing power reserves and model optimization, with capital expenditures entering a new expansion cycle [3]. - The domestic computing infrastructure is currently insufficient to meet rapidly growing demand, leading to a rise in computing power leasing as a solution [3]. Group 3: AI Terminals - The smartphone AI functionality is still under development, but there are structural innovations in hardware such as optics, foldable screens, and fingerprint recognition [4]. - The smart glasses market is gaining traction, with sales increasing, and the success of Meta & Rayban AI glasses demonstrates the potential for AI glasses to replace traditional eyewear [5]. - The transition from AI to AR is expected to enhance user experience, with optical display modules becoming a significant component in AR glasses [5]. Group 4: Investment Recommendations - Key areas to focus on include servers (Industrial Fulian, Huqin Technology), computing chips (Chipone, Cambricon, Haiguang Information), PCBs (Huidian Technology, Shenghong Technology), and power & thermal control (Heavenly Electric, Zhongheng Electric) [6]. - Other notable mentions include brands and OEMs (Xiaomi Group, Yingshi Innovation), SOCs (Lexin Technology), and storage (Zhaoyi Innovation) [6].

Summary of Huawei CloudMatrix384 Architecture and Performance Analysis Industry and Company - **Industry**: AI Infrastructure - **Company**: Huawei Core Points and Arguments 1. **Comparison with NVIDIA**: The report provides a comprehensive technical and strategic evaluation of Huawei's CloudMatrix384 AI cluster compared to NVIDIA's H100 cluster architecture, highlighting fundamental differences in design philosophy and system architecture [1][2][3] 2. **Architecture Philosophy**: Huawei's CloudMatrix384 adopts a radical, flat peer-to-peer architecture, utilizing a Unified Bus (UB) network that eliminates performance gaps between intra-node and inter-node communications, creating a tightly coupled computing entity [2][3] 3. **Performance Metrics**: The CloudMatrix-Infer service on Ascend 910C outperforms NVIDIA's H100 and H800 in terms of computational efficiency during the pre-fill and decode phases, showcasing Huawei's "system wins" strategy [3] 4. **Challenges**: Huawei faces significant challenges with its CANN software ecosystem, which lags behind NVIDIA's CUDA ecosystem in terms of maturity, developer base, and toolchain richness [3][4] 5. **Targeted Optimization**: CloudMatrix384 is not intended to be a universal replacement for NVIDIA H100 but is optimized for specific AI workloads, marking a potential bifurcation in the AI infrastructure market [4][5] Technical Insights 1. **Resource Decoupling**: The architecture is based on a disruptive design philosophy that aims to decouple key hardware resources from traditional server constraints, allowing for independent scaling of resources [6][7] 2. **Unified Bus Network**: The UB network serves as the central nervous system of CloudMatrix, providing high bandwidth and low latency, crucial for the performance of the entire system [8][10] 3. **Non-blocking Topology**: The UB network creates a non-blocking all-to-all topology, ensuring nearly consistent communication performance across nodes, which is vital for large-scale parallel computing [10][16] 4. **Core Hardware Components**: The Ascend 910C NPU is the flagship AI accelerator, designed to work closely with the CloudMatrix architecture, featuring advanced packaging technology and high memory bandwidth [12][14] 5. **Service Engine**: The CloudMatrix-Infer service engine is designed for large-scale MoE model inference, utilizing a series of optimizations that convert theoretical hardware potential into practical application performance [17][18] Optimization Techniques 1. **PDC Decoupled Architecture**: The architecture innovatively separates the inference process into three independent clusters, enhancing scheduling and load balancing [18][19] 2. **Large-scale Expert Parallelism (LEP)**: This strategy allows for extreme parallelism during the decoding phase, effectively managing communication overhead with the support of the UB network [22][23] 3. **Hybrid Parallelism for Prell**: This approach balances load during the pre-fill phase, significantly improving throughput and reducing idle NPU time [24] 4. **Caching Services**: The Elastic Memory Service (EMS) leverages all nodes' CPU memory to create a unified, decoupled memory pool, enhancing cache hit rates and overall performance [24][29] Quantization and Precision 1. **Huawei's INT8 Approach**: Huawei employs a complex, non-training-dependent INT8 quantization strategy that requires fine calibration, contrasting with NVIDIA's standardized FP8 approach [30][31] 2. **Performance Impact**: The report quantifies the contributions of various optimization techniques, highlighting the significant impact of context caching and multi-token prediction on overall performance [29][30] Conclusion - The analysis indicates that Huawei's CloudMatrix384 represents a significant shift in AI infrastructure design, focusing on specific workloads and leveraging a tightly integrated hardware-software ecosystem, while also facing challenges in software maturity and market penetration [4][5][30]