Core Viewpoint - The performance of the G100 chip from the domestic GPU startup Lishuan Technology has been called into question after its GeekBench test score of 15524, which is comparable to the Nvidia GTX 660 Ti from 13 years ago, raising concerns about its competitiveness in the market [1][2]. Group 1: Product Performance - The G100 chip, equipped with 32 compute units and a minimum core frequency of 300MHz, achieved a score of 15524 in OpenCL general computing tests, significantly lower than the RTX 4060, which scored over 100,000, making G100's performance only 6.5% of that of RTX 4060 [2]. - The G100's specifications include a mere 256MB of video memory, which is considerably below the industry standard for modern GPUs, and the actual performance may be affected by potential driver optimization issues and firmware limitations [3]. Group 2: Architecture and Design - The G100 utilizes the self-developed TrueGPU architecture, designed to meet the demands of high-performance graphics rendering and the growing need for AI applications, marking it as the first GPU architecture to integrate both capabilities [4]. - TrueGPU architecture features a unified shader and tensor engine core, which aims to enhance compatibility with AI ecosystems and provide robust computing performance, while also supporting major graphics and computing APIs such as DirectX 12.2, Vulkan 1.3, OpenGL 4.6, and OpenCL 3.0 [4].

Core Viewpoint - The article discusses the upcoming IPO of Moore Threads, a domestic GPU company in China, highlighting its rapid growth and significant backing from notable investors [1][2][3]. Company Overview - Moore Threads was founded in June 2020 and has a registered capital of 330 million yuan. The actual controller, Zhang Jianzhong, holds 44.07% of the shares [7][8]. - The company has completed six rounds of financing, raising over 4.5 billion yuan, with a valuation of 25.5 billion yuan as of last November [3][4]. Investment and Market Response - The company has attracted investments from prominent firms such as Sequoia China, Tencent, ByteDance, and Pony.ai, indicating a strong investor confidence [5]. - Following the news of its IPO progress, several related stocks, including Heertai, saw significant price increases, with Heertai rising over 8% [6]. Product and Technology Development - Moore Threads focuses on providing a full-function GPU, utilizing the MUSA architecture, which integrates AI computing acceleration, graphics rendering, video encoding/decoding, physical simulation, and scientific computing [9]. - The company has developed various products for both B-end and C-end markets, including the MTT S4000 model for large model training and the MTT X300 professional graphics card [10][13]. Competitive Landscape - Other companies in the same IPO wave include Suiyuan Technology and Birun, both focusing on AI training and computing solutions, showcasing a growing trend in the domestic chip industry [14][15].

Core Viewpoint - The progress of Moer Thread's IPO counseling status to "counseling acceptance" indicates a significant step towards its potential listing, marking a milestone for the domestic GPU chip industry in China [1][3]. Company Overview - Moer Thread, founded in 2020 by former NVIDIA executive Zhang Jianzhong, focuses on developing domestic full-function GPU chips aimed at providing computing power for AI training, graphics rendering, and scientific computing [1][2]. - The company has completed six rounds of financing, raising over 4.5 billion yuan, with investments from over 40 notable institutions including Sequoia China, Tencent, and ByteDance [1]. Product Development - Moer Thread's self-developed GPU chip utilizes the MUSA architecture, integrating four major engines: AI computing acceleration, graphics rendering, video encoding/decoding, and physical simulation, showcasing strong competitiveness [2]. - In December 2023, the company launched the fully domestic thousand-card billion model training platform, Moer Thread KUAE, highlighting its capabilities in large model training [2]. - The MTT S4000 computing card successfully passed evaluation by the China Academy of Information and Communications Technology, enhancing its market recognition [2]. Market Reaction - Following the news of Moer Thread's IPO progress, shares of its affiliated company, Heertai, surged nearly 7%, positively impacting related concept stocks [2]. - Analysts believe that if Moer Thread successfully lists, it could leverage capital market resources to expand production capacity and increase R&D investment, thereby enhancing market share [2]. Competitive Landscape - The domestic GPU market is experiencing rapid growth due to explosive AI computing demand and supportive policies for domestic alternatives, but it also faces intense competition from companies like Cambricon and Birun Technology [3]. - Despite Moer Thread's technological advancements, it still lags behind international giants like NVIDIA in terms of technology maturity and market share, posing a significant challenge for profitability post-IPO [3]. Industry Significance - The completion of Moer Thread's IPO counseling is a crucial milestone for the domestic GPU industry, with implications for both the company and the broader domestic computing chip sector [3].

Core Viewpoint - The article discusses the growing interest and procurement plans for domestic GPUs, particularly Huawei's Ascend series, in light of increasing restrictions on GPU imports from the U.S. [1] Domestic GPU Procurement Situation - ByteDance is the most proactive in adopting domestic GPUs, primarily using Cambrian and Ascend, with plans to increase procurement in 2025 [2] - Alibaba is more cautious, planning to purchase Haiguang and Cambrian cards in 2025 while continuing to stockpile NVIDIA products [3] - Tencent is the most conservative, preferring NVIDIA products, followed by H company's offerings, and only considering Cambrian and Haiguang as alternatives [3][4] Ascend 910 Series Latest Situation - As of early June, the Ascend 910B has shipped approximately 160,000 units, while the 910C has only shipped about 27,000 units due to production delays [5] - The 910C has received positive feedback for its cluster capabilities but still lags behind NVIDIA in performance [5][6] - The annual shipment target for the Ascend series has been adjusted from 850,000 to around 700,000 units due to demand fluctuations and market conditions [6] Application Scenarios and Challenges - Approximately 15% to 20% of Ascend cards are used for model training, with the majority for inference tasks [6] - High costs and technical challenges limit the widespread use of Ascend cards in high-end training applications [6][8] - The development of multi-modal models is driving higher hardware standards, with domestic manufacturers striving to catch up with international brands [7] Custom Development and Market Outlook - Collaborating with Ascend for customized development offers cost advantages but requires significant human resources for adaptation [8] - The domestic chip industry is expected to expand its market share as technology advances and market demands evolve [8]

随着美国针对GPU的限制越来越多，大家对国产GPU尤其是华为昇腾系列GPU的表现和出货量都非 常关心，前段时间老黄还说华为芯片性能已超H200，华为的CloudMatrix云架构也已经超英伟达。当 然老黄这么说肯定是有夸张的成分，他一直对美国政府出台的这些限制政策很不爽，既让英伟达损 失了很多中国的大客户，又让国产GPU在这段时间取得了突飞猛进的发展。 由于华为从未对外公布过昇腾系列的出货量，因此我们只能参考第三方的调研，本文中关于昇腾910 系列出货量的数据，是参考自本营最近的一份调研纪要。 国产GPU采购情况 相信本文的读者们应该都看到过关于CSP大厂的资本开支，字节预计2025年投入1600亿，阿里宣布 未来3年投入3800亿建设云和AI硬件基础设施，腾讯预计2025投入900亿。根据公众号"AI半导体专 研"的一个调研，在H20被禁后，国内大厂对于国产GPU卡的态度和采购计划如下： 字节最为积极，主要采用寒武纪和昇腾，并广泛测试国内其他品牌，2025年计划增加采购这些品牌 的卡。 阿里则较为谨慎，除了自研的平头哥外，还上线了部分昇腾卡，但性能表现不佳；2025年预计首次 采购海光和寒武纪的卡，同时继续 ...

Group 1 - The core point of the news is the transfer of 32.4778% equity of Shenzhen Zhongwei Electric Technology Co., Ltd. by a state-owned shareholder at a base price of 484 million yuan, reflecting a company valuation of only 1.49 billion yuan, which is notably low in the current GPU market context [1][2] - Zhongwei Electric, as a key player in GPU research and development under the China Electronics Corporation, has achieved milestones such as the successful launch and mass production of its first high-performance desktop GPU "Nanfeng No. 1" in 2022 and the completion of core IP development for "Nanfeng No. 2" in 2024 [2] - Despite these achievements, the company has faced ongoing operational losses, reporting revenue of 55.21 million yuan in 2023 with a loss of 89.05 million yuan, and a loss of 38.54 million yuan in the first ten months of 2024 [2] Group 2 - The equity transfer by China Electronics International Information Service Co., Ltd. reflects the challenges faced in GPU research and development, including the need for substantial ongoing investment and intense market competition dominated by international giants [2] - The ownership structure of Zhongwei Electric includes investments from notable firms such as Sequoia Capital and Lihua Venture Capital, raising questions about how this equity change will impact the company's future and whether new strategic investors can be found [3] - The ongoing development of the information technology industry highlights the necessity for domestic chip companies to balance technological breakthroughs with commercial sustainability [3]

Core Viewpoint - The article discusses the performance of domestic GPUs, highlighting the competitive landscape and the impact of sanctions on production capabilities. It emphasizes that while the BR100 from Birun is the most powerful, it cannot be mass-produced due to sanctions, leaving Huawei's 910C as the strongest available option [1][2]. Performance Comparison - The performance ranking of GPUs indicates that the Huawei 910C leads, followed by the Haiguang BW100, with the latter achieving approximately 87% of its potential performance. The Cambrian 590's performance is uncertain, potentially reaching 80% [2]. - The Haiguang BW100 has a deep computing capability of over 400T, which is about 50% of the H800's performance, while the 910C reaches approximately 60% of the H800's performance [1][2]. Memory and Bandwidth - Most domestic GPUs utilize HBM2e memory due to sanctions preventing the use of HBM3e. Sufficient memory is crucial for initiating inference tasks, as insufficient memory can hinder model startup [3]. - Huawei leads significantly in memory bandwidth, which is a critical factor for performance [4]. Pricing Insights - Current market prices for GPUs are as follows: Haiguang BW100 at approximately 100,000 yuan, Huawei 910B at around 70,000 yuan, and 910C at about 180,000 yuan. The Cambrian 590 has seen a price drop from 80,000-85,000 yuan to between 60,000-70,000 yuan [2]. Technical Specifications - A table summarizes key specifications of various GPUs, including memory type, bandwidth, capacity, and arithmetic intensity. For instance, the Huawei 910C features HBM2e with a bandwidth of 3.2 TB/s and a memory capacity of 64 GB [5].