以下文章来源于小鹿AI研习社 ，作者Mr. D 小鹿AI研习社 . 英伟达在H100等GPU上已实现FP8的高效支持，并通过动态缩放策略（如per-tensor scaling）和Tensor Core指 令优化，使其成为训练千亿级大模型的"标配"。然而，这些优化深度绑定英伟达硬件，国产GPU若直接照 搬，往往面临数值不稳定、训练难以收敛等问题。 FP8： 大模型时代的"算力加速器" UE8M0 FP8： 国产芯片的"妥协与突破" 小鹿AI，让你跑得更快! 我们立志传播AI前沿洞见与实践，促进AI商业应用，让积极拥抱AI的企业家取 得更大成功！ 引言 8月22日，深度求索（DeepSeek）正式发布V3.1版本大模型，技术亮点包括混合推理架构、更高的思考效率 以及更强的Agent能力。但真正引发行业热议的，是其在官微置顶中提到的"UE8M0 FP8"——这一专为下一 代国产芯片设计的浮点数格式，透露出国产AI生态正在经历一场从软件到硬件的深度协同变革。 在深度学习中，模型参数通常以浮点数（Floating Point, FP）形式存储和计算。传统的FP32（32位浮点数） 精度高但占用显存大，而FP8（8位浮 ...

Core Viewpoint - The article discusses the easing of export controls on NVIDIA's H20 GPU to China, highlighting the ongoing demand for AI GPUs in the Chinese market and the potential impact on NVIDIA's inventory and sales strategy [3][4]. Group 1: NVIDIA's H20 GPU and Market Demand - The U.S. has relaxed strict export controls on NVIDIA's H20 GPU designed for the Chinese market, which is part of a broader compromise related to China's rare earth magnet export restrictions [3]. - NVIDIA claims to have received assurances from U.S. officials to obtain necessary authorizations to resume H20 GPU sales to China [4]. - Jefferies estimates that NVIDIA currently has between 600,000 to 900,000 H20 GPUs in inventory, while the demand in China is around 1.8 million units [4]. Group 2: Capital Expenditure Projections - Jefferies has raised its forecast for AI capital expenditure in China by 40% this year to $108 billion, and increased the 2025-2030 forecast by 28% to $806 billion [5]. Group 3: Repair Market for NVIDIA GPUs in China - Due to U.S. sanctions limiting NVIDIA GPU supply, repair shops in China are thriving, focusing on older models like H100 and A100 GPUs [6][7]. - Repair costs for GPUs can reach up to $2,400, with some shops repairing around 500 chips monthly [7]. - Despite the sanctions, the demand for NVIDIA's GPUs remains high in China, as local alternatives like Huawei's products are limited [7].

行 电子 2025 年 07 月 25 日 业 研 究 行 业 动 电子 行业动态跟踪：美国发布 AI 行动计划之时，更 应重视自主可控投资机会 投资要点： 事件 1：特朗普政府发布了《美国 AI 行动计划》 7 月 23 日，特朗普政府发布了《美国 AI 行动计划》，旨在通过放宽 监管和扩大数据中心能源供应，来加速美国人工智能的发展。这一计划主 要目的之一是让美国硬件和软件成为全球 AI 创新的"标准"平台。 态 跟 踪 报告强调美国人工智能行动计划有三大支柱，分别是加速人工智能创 新、建设美国 AI 基础设施、在国际 AI 外交和安全中发挥领导作用；并在 第三支柱中表明将加强人工智能计算出口管制执行，寻求创新的出口管制 执行方法，同时堵塞现有半导体制造出口管制中的漏洞，建议在商务部的 领导下，制定关于半导体制造子系统的新出口管制。目前，美国及其盟友 对半导体制造所需的主要系统实施出口管制，但不控制许多组件子系统。 由此可以看出美国未来会进行更加严格的半导体设备和零部件出口管制， 这将进一步推进国产设备和零部件自主可控的紧迫性，利好相关产业链。 事件 2：马斯克：xAI 将在 5 年内部署等效 5000 万 ...

Investment Rating - The report does not explicitly provide an investment rating for the DCF industry in China. Core Insights - The DCF industry in China is experiencing a significant transformation driven by the increasing energy consumption and cost structure of intelligent computing centers, necessitating a shift from "equipment stacking" to "cooling-computing synergy" [5][11] - The commercial model of intelligent computing centers is evolving from traditional infrastructure hosting to a full-chain service model encompassing computing power, platforms, models, and applications, catering to diverse customer needs [5][13] - The energy consumption structure of data centers is heavily concentrated in IT equipment and cooling systems, with IT equipment accounting for 67% and cooling systems for 27% of total energy consumption, highlighting the critical role of cooling efficiency in optimizing PUE [12][26] Summary by Sections Overview of Intelligent Computing Centers - The infrastructure of intelligent computing centers is complex, involving multiple systems such as power supply, cooling, and network management, aimed at ensuring high availability and stable operation of computing devices [5][6] - The cost of intelligent computing centers is influenced by various factors including customer demand, technical solutions, redundancy design, scale, location, and equipment selection, leading to significant customization and systemic cost differences [9][10] Energy Consumption and Cost Analysis - The cost structure of intelligent computing centers is characterized by high capital investment and energy consumption, with GPU clusters like H100 representing over 85% of initial investment [11][12] - Annual operational expenditures are approaching 15% of initial costs, with energy consumption, cooling costs, and maintenance becoming dominant factors [12][11] Cooling Technology and PUE Development Trends - The cooling technology in data centers is transitioning from traditional air cooling to liquid cooling, which offers superior energy efficiency and is essential for meeting the PUE control requirements of next-generation intelligent computing centers [14][15] - Liquid cooling technologies are capable of supporting high-density heat loads, making them the mainstream cooling trend for future intelligent computing centers [15] Power Supply System Development Trends - The power supply system in data centers is evolving towards a more intelligent, redundant, and responsive multi-level energy quality collaborative architecture, with UPS systems playing a central role in energy governance and system coupling [16][18] Regional Distribution of Intelligent Computing Centers - The layout of intelligent computing centers in China shows a pattern of concentration in the eastern regions, driven by high demand for computing power and robust infrastructure, while the central and western regions are developing to support resource allocation [19][20] Participants in the Intelligent Computing Center Market - The projects in the intelligent computing center sector are primarily initiated by local governments, while major internet and cloud companies are consolidating computing resources, indicating a dual structure of policy guidance and market concentration [21][22] Data Center Power Demand - The power demand for data centers is expected to grow significantly, with China's core IT load electricity demand projected to exceed 285 TWh by 2026, positioning China as a key player in global power-consumption growth [23][24] Average Annual PUE of Data Centers - The average PUE of global and Chinese data centers is currently stagnating, indicating a need for structural innovation in cooling systems and energy architecture to enhance energy efficiency [25][26]

Core Insights - Tesla has expanded its artificial intelligence training computing capabilities at its Texas Gigafactory by adding an additional 16,000 H200 GPUs, bringing the total number of H100 GPUs used in the Cortex supercomputing cluster to 67,000 [1] Group 1 - The addition of 16,000 H200 GPUs enhances Tesla's AI training capabilities [1] - The total number of H100 GPUs in use is now equivalent to 67,000 [1]

Group 1 - The core point of the articles highlights the increasing competition among tech giants for top AI talent, with a focus on salary and computational resources as key factors for recruitment [1][2] - Meta CEO Mark Zuckerberg emphasizes that AI researchers are now more interested in having fewer direct reports and access to maximum GPU resources, which are crucial for building and training AI models [1] - The demand for NVIDIA's H100 GPUs has surged, as they are considered essential for AI research, and companies are willing to offer substantial compensation packages to attract talent [1][2] Group 2 - Other AI companies, like Perplexity, confirm the trend, stating that significant computational resources are necessary to attract top talent, with one researcher reportedly refusing an offer due to insufficient GPU availability [2] - Meta is actively recruiting from competitors such as Google, Anthropic, and OpenAI, offering signing bonuses that can reach up to $100 million [2]

Don't just invest—dominate with Tech Contrarians' realized return on closed positions of 65.8% since inception. You'll get exclusive insights into high-focus stocks, curated watchlists, one-on-one portfolio consultations, and everything from live portfolio tracking to earnings updates on 50+ companies. Subscribe today for 20% off. Analyst's Disclosure:I/we have no stock, option or similar derivative position in any of the companies mentioned, and no plans to initiate any such positions within the next 72 ho ...

Core Viewpoint - Intel's CEO, Pat Gelsinger, stated that the company is "too late" in catching up in the AI training sector, acknowledging Nvidia's strong market position [3] Group 1: AI Market Position - Intel is shifting its focus from AI training to inference, particularly in edge computing and agentic AI, as predictions suggest the inference market will eventually surpass the training market [3] - The current AI training data centers are dominated by Nvidia (H100) and AMD (MI300X) GPUs, with major cloud operators like Google, Amazon, and Microsoft developing their own AI chips [3] Group 2: Company Restructuring - Intel is undergoing a restructuring process, which includes significant layoffs, with reports indicating up to 2,392 layoffs in Oregon and around 4,000 in other states [4] - The layoffs will affect various positions, including hundreds of technical staff and engineers, and represent about 20% of Intel's workforce in Oregon [4] - Following the layoffs, Intel's workforce will decrease by approximately 16,000, with a projected market value of $102 billion by July 2025 [4]

Nvidia (NVDA -0.85%) stock has soared by 870% since the start of 2023, catapulting its market capitalization to a whopping $3.5 trillion. Demand continues to exceed supply for the company's graphics processing units (GPUs) for the data center, which are the most powerful chips in the world for developing artificial intelligence (AI) models.Nvidia CEO Jensen Huang says new AI reasoning models, which spend more time thinking to produce accurate responses, require up to 1,000 times more computing capacity than ...

Investment Rating - The semiconductor production equipment industry is rated as Attractive [5][70]. Core Insights - The imminent launch of DeepSeek R2, which features 1.2 trillion parameters and significant cost efficiencies, is expected to positively impact the Japanese semiconductor production equipment (SPE) industry [3][7][11]. - The R2 model's capabilities include enhanced multilingual support, broader reinforcement learning, multi-modal functionalities, and improved inference-time scaling, which could democratize access to high-performance AI models [7][9][11]. - The development of efficient AI models like R2 is anticipated to increase demand for AI-related SPE, benefiting companies such as DISCO and Advantest [11]. Summary by Sections DeepSeek R2 Launch - DeepSeek's R2 model is reported to have 1.2 trillion parameters, a significant increase from R1's 671 billion parameters, and utilizes a hybrid Mixture-of-Experts architecture [3][7]. - The R2 model offers cost efficiencies with input costs at $0.07 per million tokens and output costs at $0.27 per million tokens, compared to R1's $0.15-0.16 and $2.19 respectively [3][7]. Industry Implications - The launch of R2 is expected to broaden the use of generative AI, leading to increased demand for AI-related SPE across the supply chain, including devices like dicers, grinders, and testers [11]. - The report reiterates an Overweight rating on DISCO and Advantest, which are positioned to benefit from the anticipated increase in demand for AI-related devices [11]. Company Ratings - DISCO (6146.T) is rated Overweight with a target P/E of 25.1x [12]. - Advantest (6857.T) is also rated Overweight, with a target P/E of 14.0x [15].