在正在举办的半导体行业会议 Hot Chips 2025 上，TogetherAI 首席科学家 Tri Dao 公布了 FlashAttention-4 。 机器之心报道 编辑：Panda 在这个新版本的 FlashAttention 中，Tri Dao 团队实现了两项关键的算法改进。 一、它使用了一种新的在线 softmax 算法，可跳过了 90% 的输出 rescaling。 二、为了更好地将 softmax 计算与张量核计算重叠，它使用了指数 (MUFU.EX2) 的软件模拟来提高吞吐量。 此外，FlashAttention-4 使用的是 CUTLASS CuTe Python DSL，其移植到 ROCm HIP 的难度要高出 10 倍，而 CUDA C++ 移植到 ROCm HIP 则更容易。 据介绍，在 Backwell 上，FlashAttention-4 的速度比英伟达 cuDNN 库中的注意力核实现快可达 22%！ 有意思的是，Tri Dao 还宣布，在执行 A@B+C 计算时，对于 Blackwell 上在归约维度 K 较小的计算场景中，他使用 CUTLASS CuTe-DSL 编写的核（k ...

Nvidia NVDA stock is trading higher on Monday as it rolled out its Jetson AGX Thor developer kit and production modules, a next-generation robotics platform built to power millions of robots, including humanoids, with advanced AI capabilities.The Blackwell GPU–powered system offers record-breaking performance and efficiency, reinforcing Nvidia’s dominance in AI infrastructure as analysts project stronger earnings driven by its GB200 and Blackwell product ramps.The Blackwell GPU–powered platform delivers up ...

(责任编辑：宋政 HN002) 【免责声明】本文仅代表作者本人观点，与和讯网无关。和讯网站对文中陈述、观点判断保持中立，不对所包含内容 的准确性、可靠性或完整性提供任何明示或暗示的保证。请读者仅作参考，并请自行承担全部责任。邮箱： news_center@staff.hexun.com 格隆汇8月25日｜据集微网，英伟达已向美国政府提交了一款B30芯片，以获得向中国出口的许可。相 关谈判于今年早些时候开始，该芯片的峰值性能仅为标准Blackwell GPU性能的80%。美国总统特朗普 曾表示，如果英伟达的Blackwell处理器性能比其顶级产品低至少30%，他将允许其上市。特朗普表 示："我可能会就一款'略微增强性能'的Blackwell处理器达成协议。换句话说，性能要降低30%到 50%。"相比之下，英伟达HGX H20的性能仅为全功能版H100的50%左右，尤其是在多GPU配置下。 ...

Core Viewpoint - Nvidia is currently overvalued in a market that is not performing optimally, with high expectations set for its upcoming fiscal second-quarter results [1][3][11] Group 1: Nvidia's Market Position and Performance - Nvidia has seen an approximately 1,100% increase in its stock price since the beginning of 2023, indicating strong performance [3] - The company has established itself as a leader in AI-graphics processing units (GPUs), with its Hopper (H100) and Blackwell GPUs being widely deployed in high-compute data centers [5] - Nvidia's gross margin reached a high of 78.4% during the first quarter of fiscal 2025, driven by a backlog for its AI-GPUs allowing it to command premium pricing [6] Group 2: Competitive Landscape - Nvidia faces increasing competition from Advanced Micro Devices and Huawei, which are ramping up production of their own data-center chips [7] - Major customers of Nvidia are developing their own AI GPUs, which, while not as powerful, are cheaper and not backlogged, potentially impacting Nvidia's market share [9][10] Group 3: Valuation Concerns - Nvidia's trailing-12-month price-to-sales (P/S) ratio was above 30, indicating a valuation that may not be sustainable historically [12][13] - The overall market is experiencing high valuations, with Nvidia contributing to the S&P 500's elevated Shiller price-to-earnings (P/E) ratio [14][15] Group 4: Historical Context and Future Outlook - Historical trends suggest that companies leading new technological innovations often face valuation corrections after initial hype [18][21] - Despite impressive demand for AI infrastructure, many businesses are not yet optimizing their AI solutions, indicating that the market may have overestimated the immediate impact of AI [20]

Core Insights - Nebius reported a significant year-over-year sales increase of 625% and a quarter-over-quarter growth of 106% in Q2 2025, with core AI infrastructure business EBITDA turning profitable ahead of management expectations [3][4] - The company raised its annual recurring revenue (ARR) guidance to $900 million - $1.1 billion, up from a previous range of $750 million - $1 billion, reflecting an 8.57% increase in midpoint guidance [5][6] - Nebius plans to achieve 220 MW of connected power by the end of FY 2025 and 1 GW by the end of FY 2026, with a capital expenditure of $2 billion for 2025 [7] Financial Performance - The adjusted EBITDA loss for the group improved to $21 million, a year-over-year improvement of $31 million [4] - The company’s revenue guidance for core business remains unchanged at $400 million - $600 million, while the overall group revenue guidance is set at $450 million - $630 million [6] - Analysts have raised revenue growth estimates for FY 2025 and FY 2026 to 382.9% and 158.23%, respectively [6] Market Position and Valuation - Nebius was recently included in the Wedbush IVES AI 30 Index, which may add further premium to its already high valuation [8] - The projected price-to-sales ratio for FY 2026 is 11.6x, with potential for it to reach 15x by the end of 2026, suggesting a market cap of approximately $22.05 billion in the next 12-18 months [8] - Despite positive performance, the stock is considered crowded, and further accumulation may not be advisable at this time [9]

Core Insights - The article discusses the rapid development of artificial intelligence (AI) technology and its potential investment opportunities in the market [2]. Group 1: OpenAI Developments - OpenAI has launched a new subscription plan in India for under $5 per month, aimed at expanding its AI market services, allowing users to generate more images and interact more frequently with the chatbot compared to the free version [4]. Group 2: Arm Holdings and Chip Development - Arm Holdings has hired Amazon's AI chip director Rami Sinno to participate in its autonomous chip development plan, focusing on creating chips for large AI applications [4]. - Arm's business model primarily involves designing core architectures and licensing them to clients, with significant market presence in smartphones and data center chips [4]. Group 3: Nvidia's Stock Performance and Analyst Predictions - Nvidia's stock has risen over 30% this year, with analysts raising target prices due to the insatiable demand for AI and revenue opportunities from the Chinese market [5][6]. - Analysts expect Nvidia's Q2 revenue to be around $458 billion, with earnings per share (EPS) projected at $1.00, driven by the demand for AI computing [6]. - Cantor Fitzgerald raised its target price for Nvidia from $200 to $240, citing endless demand for AI computing and increased capital expenditures from large tech companies [6][7]. - Mizuho analysts noted a rise in capital expenditure expectations from 38% to 54% year-over-year, predicting Nvidia's Q2 revenue at $462 billion and EPS at $1.01 [6]. Group 4: Nvidia's Future Earnings Expectations - Analysts predict Nvidia's future earnings will exceed expectations, with Q2 revenue estimates ranging from $466 billion to $480 billion and EPS estimates from $1.03 to $1.06 [7]. - The growing demand for inference, or generating new content based on real data, is a key factor driving enthusiasm for Nvidia's stock [7]. Group 5: OpenAI's Market Position - OpenAI's CEO Sam Altman acknowledged the existence of a market bubble around AI but emphasized the technology's importance and lasting impact [12][13]. - OpenAI aims to surpass Meta's platforms in user engagement, currently boasting over 700 million weekly users [13].

Group 1 - Meta Platforms plans to restructure its AI business into four departments, including a new lab called TBD Lab, within the next six months [4] - OpenAI's CEO Sam Altman aims to invest "trillions of dollars" in AI infrastructure, believing that society will not regret such investments in the long term [4] - WeRide received a multi-million dollar investment from Grab to accelerate the deployment of L4 Robotaxis in Southeast Asia [4] Group 2 - NVIDIA's latest 13F filing reveals that as of June 30, 91.36% of its public holdings are concentrated in AI cloud computing service provider CoreWeave, with a total investment of $3.96 billion [5][6] - CoreWeave's Q2 revenue reached $1.2 billion, a year-over-year increase of over 300%, although its stock price recently fell nearly 21% due to lower-than-expected revenue growth and plans for significant capital expenditures [6] - Analysts predict that CoreWeave's revenue could grow by 127% next year, potentially reaching $11 billion, highlighting NVIDIA's confidence in the AI infrastructure sector [7] Group 3 - Morgan Stanley has raised its iPhone production forecast for September by 8%, now estimating 54 million units, citing better-than-expected sales in June [11][12] - The positive revision is attributed to a reduction in iPhone channel inventory below normal levels, creating greater opportunities for channel filling in September [12] - Analysts expect that the production of the iPhone 17 will remain stable at 80 to 85 million units by the second half of 2025, with a slight year-over-year decline [13]

Core Insights - Data centers are becoming the core engine driving global economic and social development, marking a new era for the semiconductor industry, driven by AI, cloud computing, and large-scale infrastructure [2] - The demand for chips in data centers is evolving from simple processors and memory to a complex ecosystem encompassing computing, storage, interconnect, and power supply [2] AI Surge: The Arms Race in Data Centers - The explosion of artificial intelligence, particularly generative AI, is the strongest catalyst for this transformation, with AI-related capital expenditures surpassing non-AI spending, accounting for nearly 75% of data center investments [4] - By 2025, AI-related investments are expected to exceed $450 billion, with AI servers rapidly increasing from a few percent of total computing servers in 2020 to over 10% by 2024 [4] - Major tech giants are engaged in a fierce "computing power arms race," with companies like Microsoft, Google, and Meta investing hundreds of billions annually [4] - The data center semiconductor market is projected to expand significantly, reaching $493 billion by 2030, with data center semiconductors expected to account for over 50% of the total semiconductor market [4] Chip Dynamics: GPU and ASIC Race - GPUs will continue to dominate due to the increasing complexity and processing demands of AI workloads, with NVIDIA transforming from a traditional chip designer to a full-stack AI and data center solution provider [7] - Major cloud service providers are developing their own AI acceleration chips to compete with NVIDIA, intensifying competition in the AI chip sector [7] - High Bandwidth Memory (HBM) is becoming essential for AI and high-performance computing servers, with the HBM market expected to reach $3.816 billion by 2025, growing at a CAGR of 68.2% from 2025 to 2033 [8] Disruptive Technologies: Redefining Data Center Performance - Silicon photonics and Co-Packaged Optics (CPO) are key technologies addressing high-speed, low-power interconnect challenges in data centers [10] - The adoption of advanced packaging technologies, such as 3D stacking and chiplets, allows semiconductor manufacturers to create more powerful and flexible heterogeneous computing platforms [12] - The shift to direct current (DC) power supply is becoming essential due to the rising power density demands of modern AI workloads, with power requirements for AI racks expected to reach 50 kW by 2027 [13] Cooling Solutions: Liquid Cooling Technology - Liquid cooling technology is becoming a necessity for modern data centers, with the market projected to grow at a CAGR of 14%, exceeding $61 billion by 2029 [14] - Various types of liquid cooling methods, including Direct Chip Liquid Cooling (DTC) and immersion cooling, are being adopted to manage the heat generated by high-performance AI chips [15] - Advanced thermal management strategies, including software-driven dynamic thermal management and AI model optimization, are crucial for maximizing future data center efficiency [16] Future Outlook - The future of data centers will be characterized by increasing heterogeneity, specialization, and energy efficiency, with chip design evolving beyond traditional CPU/GPU categories [17] - Advanced packaging technologies and efficient power supply systems will play a critical role in shaping the next generation of green and intelligent data centers [17]

Core Viewpoint - Nvidia has heavily invested in AI cloud computing, with 91.36% of its public holdings concentrated in CoreWeave, totaling $3.96 billion [1][3][4] Group 1: Investment Strategy - Nvidia's investment strategy includes a diversified portfolio in AI-related companies, with significant stakes in Arm, Applied Digital, Nebius, and Recursion Pharmaceuticals [3][8] - CoreWeave is positioned as a key asset in Nvidia's portfolio, reflecting confidence in AI infrastructure [6][10] Group 2: CoreWeave's Performance - CoreWeave reported Q2 revenue of $1.2 billion, a year-over-year increase of over 300%, although its stock price recently dropped nearly 21% due to lower-than-expected revenue growth and increased capital expenditure plans [4][11] - Analysts predict CoreWeave's revenue could grow by 127% next year, potentially reaching $11 billion, highlighting strong demand for AI computing [6] Group 3: Challenges and Market Dynamics - CoreWeave faces challenges, including over $11 billion in total debt and cash consumption during network expansion, alongside a significant stock unlock event that may pressure its share price [10][11] - The upcoming stock unlock will test market enthusiasm for AI infrastructure, as approximately 84% of Class A shares will be released, primarily held by insiders and Nvidia [4][11]

Core Insights - Nvidia has heavily invested in CoreWeave, allocating 91.36% of its public holdings, amounting to $3.96 billion, indicating a strong belief in AI cloud computing services [1][2] - CoreWeave's revenue for Q2 reached $1.2 billion, showing a year-over-year growth of over 300%, although its stock price recently dropped nearly 21% due to lower-than-expected revenue growth and increased capital expenditure plans [1][4] - The company is facing a significant unlock of approximately 84% of Class A shares, which may lead to market volatility [1][4] Investment Diversification - Besides CoreWeave, Nvidia holds stakes in several other AI-related companies, including Arm (4.11%), Applied Digital (1.79%), Nebius (1.52%), and Recursion Pharmaceuticals (0.90%) [3] - Arm is expanding its energy-efficient architecture into data centers, while Applied Digital focuses on high-performance computing infrastructure [3] - Recursion Pharmaceuticals applies machine learning in drug discovery, showcasing Nvidia's interest in AI-driven healthcare [3] Challenges Facing CoreWeave - CoreWeave is burdened with over $11 billion in total debt and is consuming cash while expanding its network [4] - Despite these challenges, the company has a strong customer base, with major tech firms like Microsoft and Meta increasing their capital expenditure forecasts [4] - CoreWeave has secured a $4 billion four-year expansion agreement with OpenAI and announced a $9 billion stock acquisition of Core Scientific [4]