AI大模型训练
Search documents
研判2025!中国数据交换系统行业概述、市场现状及发展趋势分析:数据交换系统正从传统网络连接设备向算力基础设施核心枢纽跃迁,规模不断增长[图]
Chan Ye Xin Xi Wang· 2026-01-08 01:27
内容概况:中国数据交换系统行业正迎来一轮由多重战略级需求强力驱动的结构性增长。2024年,中国 数据交换系统行业市场规模约为137.49亿元,同比增长7.05%。随着"东数西算"国家战略、AI大模型训 练及5G边缘计算规模化部署三重动能的叠加释放,推动数据交换系统从传统网络连接设备向算力基础 设施核心枢纽跃迁。 相关上市企业:新晨科技(300542)、三维天地(301159) 相关企业:中芯国际集成电路制造有限公司、紫光国芯微电子股份有限公司、兆易创新科技股份有限公 司、北方华创科技集团股份有限公司、武汉光迅科技股份有限公司、中际旭创股份有限公司、广东风华 高新科技股份有限公司、立讯精密工业股份有限公司、沪士电子股份有限公司、浪潮电子信息产业股份 有限公司、曙光信息产业股份有限公司、北京光环新网科技股份有限公司 关键词:数据交换系统、数据交换系统市场规模、数据交换系统行业现状、数据交换系统发展趋势 数据交换系统采用双主机架构,由部署于高安全网络的TAS(信任端)与低安全网络的UAS(非信任 端)协同组成:TAS作为安全门户,负责数据安全检查、内容过滤、病毒查杀、访问控制及协议转换, 集成ETL框架实现异构数据 ...
中国新型芯片突破光刻机限制,北大团队研发高精度模拟矩阵计算芯片
Xin Lang Cai Jing· 2025-12-31 02:24
#中国新型芯片绕开光刻机限制# 中国新型芯片问世了! 北京大学孙仲教授团队研究出的这个基于阻变存储器的高精度、可扩展模拟矩阵计算芯片,意义很大, 说白了就是换了个路子做计算,来绕开光刻机的限制。 以前模拟计算因为"算不准"而被淘汰,现在他们一举解决了这个老大难问题。这样一来,芯片做矩阵计 算时,速度更快、更省电,而且用28纳米这种成熟工艺就能造,不用去挤高端光刻机的独木桥。 眼下芯片还在实验阶段,适合中等规模的任务,比如应用于6G和机器人等,他们还计划两年内把芯片 规模做大,先在这些领域做起来。 长远看,它是为AI大模型训练、超算这些"算力吞金兽"准备的一个高能效后备方案。等这些需求爆发 时,咱们手里就有现成的技术牌可以打了。 2017 11 专访 北大孙仲团队 支撑6G和AI训练,可在28纳米 及以上成熟工艺量产 ...
绕开光刻机“卡脖子” 中国新型芯片问世!专访北大孙仲:支撑AI训练和具身智能 可在28纳米及以上成熟工艺量产
Mei Ri Jing Ji Xin Wen· 2025-12-30 00:36
当AI时代算力集群规模正逐步从万卡向十万卡、百万卡甚至千万卡升级时,一支中国团队悄然另辟蹊径。 今年10月,北京大学人工智能研究院/集成电路学院双聘助理教授孙仲与北京大学集成电路学院蔡一茂教授、王宗巍助理教授率领的团队成功研制出基于 阻变存储器的高精度、可扩展模拟矩阵计算芯片,在全球范围内首次将模拟计算的精度提升至24位定点精度,让未来同等任务下使用更少的计算卡成为可 能。 这是一种完全不同于目前所有商用量产芯片的新型芯片,计算精度从1%跃升至千万分之一;可以支撑6G、具身智能及AI大模型训练等多个前沿场景;更 重要的是,它可在28纳米及以上成熟工艺量产,绕开光刻机"卡脖子"环节。 事实上,AI大模型、具身智能、6G等应用背后都是矩阵计算,AI推理是做矩阵乘法,AI训练是在解矩阵方程。而英伟达的崛起正是得益于GPU(图形处 理器)很擅长做矩阵计算。 但如今摩尔定律趋于终结,晶体管很难再微缩,所以业界现在只能横向堆计算卡:少则百卡,多则万卡、十万卡。但这样的方式我认为是不可持续的—— 能耗、碳排放均呈指数级上升,与国家"双碳"目标相悖。因此,我认为需要探索一种不同的计算范式,即模拟(类比)计算。 模拟计算并非全 ...
华福证券:AI大模型训练进入“万卡集群”时代 OCS交换机有望步入发展快车道
Zhi Tong Cai Jing· 2025-12-23 09:13
华福证券发布研报称,随着AI大模型训练进入"万卡集群"时代,传统基于电交换的网络架构正面临网络 功耗高、带宽要求高的瓶颈。OCS交换机的"全光交换"机制从根本上保证网络的极致带宽和前瞻兼容 性。目前商用OCS交换机主流为MEMS技术方案,占比超过70%。核心环节上,国内涌现出一批优秀的 OCS核心零部件供应商,业务涵盖MEMS振镜、透镜、光器件等核心环节。 MEMS方案是目前市场的主流,占比超过70%,尤其在谷歌等企业中被广泛采用。谷歌自2022年起在其 传统数据中心中导入该技术,并于2023年应用于TPU V4,计划在2025年的TPU V5P中进一步推广。根 据QY Research,2024年MEMS技术方案的OCS交换机市场规模为192.36百万美元,约占整个全光交换 (OCS)交换机市场52.49%的份额,预计2031年将达到836.79百万美元,届时全球占比将减少至41.38%。 国产厂商积极参与OCS交换机研发与制造 整机方面:华为全联接大会2024上,华为重磅推出了其最新的全光交换机——Huawei OptiXtrans DC808;凌云光参与OCS整机研发,压电陶瓷路线代表;另外中际旭创(3 ...
浩云长盛彭轶峰:兼容并蓄和系统性思维是应对IDC行业急速迭代的生存法则
Huan Qiu Wang Zi Xun· 2025-12-18 06:31
【环球网科技报道 记者 林迪】当前,数据中心(IDC)正经历着前所未有的深刻变革。它不再仅仅是 数据存储与计算的物理空间,更被赋予了"算力中心"的全新使命,成为支撑AI大模型训练、算法迭代与 生态演进的关键基础设施。如何在这场技术革命中保持领先?IDC行业面临着怎样的挑战与机遇? 来源:环球网 近日,在第二十届中国IDC产业年度大典(IDCC2025)暨数字基础设施科技展(DITExpo)上,浩云长 盛集团技术研发AVP彭轶峰认为核心关键词是"兼容"与"前瞻"。 "今天很多数据中心被淘汰,本质上是因为过去十几年,二十年的路径选择而引起的,因为它不是一个 按照未来的发展去重新建设的路径。"彭轶峰一针见血地指出,"很早以前路径是从改造开始的,也就是 说你的改造就天然的决定了只能用眼前所需要的标准去设计数据中心。"这种源于"改造"的思维惯性, 导致许多从业者的意识形态和认知停留在了起点,一切创新都以当下的标准为蓝本进行想象。 然而,AI时代的需求是动态且不可预知的。彭轶峰强调:"AI的特点就是我们看不到它的全貌,所以它 的不确定性,决定了未来AI的发展其实是有一定的弹性、方向差异和不可预知性。面对如此充满变数 的未 ...
四川1ms城市算网加速成型 重塑算力运载新格局
Huan Qiu Wang Zi Xun· 2025-12-11 06:24
Core Insights - The article emphasizes the importance of computing power as a core driver for urban smart transformation and industrial upgrading in the context of the booming digital economy [1] - The implementation of the "1ms Urban Computing Network" is highlighted as a key initiative to enhance efficient computing power transportation, aligning with national strategies and reshaping urban development models [1][2] Group 1: Development Plans and Strategies - The "High-Quality Development Action Plan for Computing Power Infrastructure" aims to improve the efficient transportation capacity of computing power, with latency requirements set at no more than 1ms for major urban computing infrastructures and 5ms for national hub nodes [1] - Sichuan province is leveraging the "East Data West Computing" national strategy to break regional computing barriers and transform computing resources into public services for cities [1][2] Group 2: Technological Innovations - Sichuan's 1ms Urban Computing Network integrates advanced technologies such as all-optical switching (OXC), 400G transmission, and intelligent scheduling to achieve direct access to computing resources [4] - The network's architecture includes a three-tier structure (core, inner ring, outer ring) to create a unified computing network across Chengdu and 20 other cities, achieving latency under 1ms for major data centers [6] Group 3: Role of Telecom Operators - Major telecom operators are leading the construction of the 1ms Urban Computing Network, utilizing their extensive network resources and technical capabilities to ensure comprehensive coverage and efficient scheduling of computing resources [5] - Sichuan Telecom plans to invest approximately 180 million yuan to build 300 PFLOPS of domestic computing power by 2025, with a goal to exceed 1 EFLOPS by 2026 [7] Group 4: Industry Applications and Impact - The 1ms Urban Computing Network significantly empowers various industries, enabling real-time data collection, analysis, and feedback in smart manufacturing scenarios [9] - The network supports advanced manufacturing processes, as demonstrated by the success of companies like Lynk & Co, which utilizes low-latency networks for quality management and production efficiency [11][12] Group 5: Future Directions - Despite the achievements, there is a call for continued efforts to build a low-latency, high-reliability, and secure computing network development system [14] - Future initiatives will focus on enhancing service capabilities, promoting technological innovations, and fostering collaborative ecosystems among government, enterprises, academia, and research institutions [14]
华安证券:OCS光交换机有望迎来高速成长期 建议关注赛微电子等
Zhi Tong Cai Jing· 2025-12-08 07:49
Core Insights - The training of AI large models requires high communication bandwidth, low latency, and low power consumption, making Optical Circuit Switches (OCS) an ideal interconnection solution due to their high bandwidth and low latency characteristics [1][3] - The global OCS optical switch market is projected to grow from $0.07 billion in 2020 to $0.78 billion by 2025, with a compound annual growth rate (CAGR) of 62%, and is expected to reach $2.02 billion by 2031, with a CAGR of approximately 17.2% from 2025 to 2031 [3] Industry Overview - OCS is primarily applied in three major scenarios for AI computing clusters: Scale-Up (enhancing single-node performance), Scale-Out (multi-node collaboration), and Scale-Across (interconnecting across data centers) [2] - In a Google TPU cluster, a setup with 4,096 TPU v4 chips requires 48 OCS switches, indicating a TPU to OCS ratio of approximately 85:1, which improves to 192:1 with the future TPU v7 cluster [2] Market Dynamics - The OCS market is currently concentrated, with the top four manufacturers expected to hold about 69% of the market share by 2025, with Google and Coherent being key players [3] - The OCS industry chain consists of upstream core components, midstream equipment integration, and downstream applications, with high technical barriers and significant value concentration in upstream core components like MEMS micro-mirror arrays [3] Company Focus - Ying Tang Intelligent Control (300131) is expanding from electronic component distribution to semiconductor design and manufacturing, planning to acquire Guilin Guanglong Integration by 2025 to strengthen its OCS full-process layout [4] - Saiwei Electronics (300456) is a leader in MEMS process development and wafer manufacturing, with a significant portion of its revenue (83%) coming from MEMS business, and is expected to benefit from the growing demand for AI computing [4]
未来网络试验设施正式投入运行,完成120项重大创新试验
Huan Qiu Wang Zi Xun· 2025-12-06 01:50
Core Insights - The Future Network Experimental Facility, China's first major national technology infrastructure in the information and communication sector, has officially commenced operations [1] Group 1: Facility Overview - The facility is located in Nanjing, Jiangsu, and was completed in August 2024 [1] - It covers 40 cities nationwide, featuring 88 backbone network nodes and 133 edge network nodes, with a total optical transmission length exceeding 55,000 kilometers [1] - The facility can support 4,096 heterogeneous services for parallel testing and is capable of interconnecting with existing domestic and international networks [1] Group 2: Performance Metrics - The facility enables efficient, high-speed, low-latency, and low-jitter data transmission, with a packet loss rate of only one in a million [1] Group 3: Service and Collaboration - To date, the facility has served major national research institutions such as the National Astronomical Observatory and the Institute of High Energy Physics, as well as telecom operators like China Telecom, China Mobile, China Unicom, and China Broadcasting Network [1] - It has collaborated with universities including Peking University, Nanjing University, Zhejiang University, and the Chinese University of Hong Kong, along with leading companies like Huawei, H3C, and Baidu, completing 120 significant innovation experiments [1] - The experiments cover critical dimensions such as core chips, network operating systems, routing control, security and trust, large-scale networking, and new AI services [1]
存储产业“换挡”提速,DDR5普及进入快车道
Huan Qiu Shi Bao· 2025-11-24 03:23
Core Insights - The storage market is experiencing a significant structural change, with DDR4 memory chip prices rising and surpassing DDR5, indicating the beginning of the transition to the DDR5 era [1][3] - The price increase of DDR4 is attributed to supply-demand imbalances caused by leading companies like Samsung and SK Hynix reducing DDR4 production in favor of DDR5, leading to a price gap that is expected to be a temporary fluctuation [3][4] - The demand for higher data processing capabilities in applications such as AI and cloud gaming is driving the transition to DDR5, which offers superior bandwidth and performance compared to DDR4 [3][4][5] Market Dynamics - Since June, the price of 16GB DDR4 chips has increased from $7.01 to $8.59, while DDR5 prices rose from $5.85 to $6.17, with the price difference continuing to widen [3] - The shift towards DDR5 is not only a hardware upgrade but also a strategic move for financial institutions to enhance their core competitiveness in high-performance digital environments [4] Technological Advancements - DDR5 represents a generational leap over DDR4, with double the bandwidth, improved capacity and energy efficiency, and integrated ECC error correction, significantly reducing the risk of data center downtime [4] - The push for DDR5 adoption aligns with national development goals for high-level technological self-reliance and is seen as a critical factor in supporting the digital economy [4][5]
创业板50指数上涨0.88%,光模块和电池板块表现强劲
Xin Lang Cai Jing· 2025-11-10 11:41
Market Overview - The A-share market showed an overall upward trend last week, with major indices recording positive growth. The CSI 300 index rose by 0.82%, while the CSI 500 index slightly decreased by 0.04%. The ChiNext 50 index performed particularly well, increasing by 0.88% [1] - The average daily trading volume in the A-share market remained around 2 trillion yuan, indicating an increase in market activity [1] Industry Highlights - Key sectors attracting market attention include photovoltaic, new energy, and cyclical industries such as coal, steel, and chemicals. Investors are advised to focus on new energy and photovoltaic sectors, utilizing ETF products like the ChiNext 50 ETF, which has 38% exposure to new energy [1] - The ChiNext 50 index reported a year-on-year growth rate of 49% in net profit attributable to shareholders for Q3 2025, alleviating valuation pressure and enhancing investment value [1] - The ChiNext serves as a direct financing platform for innovative enterprises, supporting the development of "three innovations" (innovation, creation, and creativity) and "four new" (new technologies, new industries, new business formats, and new models) [1] Sector Performance - In the optical module and battery sectors, the ChiNext 50 index outperformed the ChiNext index and other mainstream indices. Despite a capital outflow in the optical module sector last week, long-term demand remains strong, driven by the need for 800G/1.6T optical modules due to AI model training [2] - Major cloud providers in North America, including Microsoft, Google, Meta, and Amazon, increased their capital expenditures to a total of $96.4 billion in Q3 2023, a 68% increase year-on-year. The demand for 1.6T optical modules is expected to be revised up to 20 million units by 2026 [2] - The photovoltaic sector rebounded significantly last week, supported by policy guidance from the Ministry of Industry and Information Technology, which emphasized industry self-discipline to promote the coordinated development of photovoltaic and energy storage [2] - The global photovoltaic installation capacity is projected to exceed 500 GW by 2025, providing strong support for the industry's long-term development [2] Pharmaceutical Sector - The pharmaceutical and biotechnology sector experienced a decline last week and is currently undergoing a technical adjustment. The results of medical insurance negotiations indicate that 127 drugs outside the catalog will participate in negotiations, presenting opportunities for some innovative drugs [3] - The increase in flu cases in northern regions poses challenges for related companies. However, some CXO companies reported over 40% year-on-year revenue growth in Q3 2023, demonstrating strong market competitiveness [3] - Long-term innovation remains a key driver in the pharmaceutical industry, with accelerated global licensing of new therapies such as ADC and bispecific antibodies, and significant potential for domestic companies' internationalization [3] ChiNext 50 ETF - The ChiNext 50 ETF (code: 159949) tracks the ChiNext 50 index and adopts standards focused on the "three innovations" and "four new" criteria, primarily selecting leading companies in five major technology sectors: new energy vehicles, biomedicine, electronics, photovoltaics, and internet finance [3] - The index reflects the overall performance of 50 high liquidity and market capitalization companies on the ChiNext, showcasing high investment value. The ChiNext 50 ETF has a strong liquidity profile, with an average daily trading volume of 1.497 billion yuan over the past year, ranking among the top ETFs on the Shenzhen Stock Exchange [3] - The latest fund size of the ChiNext 50 ETF is 26.974 billion yuan, making it one of the larger funds related to the ChiNext market [3]