Core Insights - Shanghai Xinen Jiyuan Technology Co., Ltd. has been established with a registered capital of 5 million RMB, fully owned by Shanghai Xinhong Intelligent Technology Co., Ltd. [1] Company Overview - The legal representative of Shanghai Xinen Jiyuan Technology Co., Ltd. is Wu Minhua [1] - The company is classified as a limited liability company with a natural person investment or holding [1] - The business scope includes technology services, development, consulting, and various applications in artificial intelligence and 5G communication technology [1] Business Activities - The company engages in software development, information technology consulting, and network security software development [1] - It also covers a wide range of sales activities, including household appliances, electronic products, and machinery [1] - The company is authorized to conduct construction engineering projects and value-added telecommunications services, subject to approval [1]

HUAWEI X HUXIU 没有人不在期待大模型能够成为下一个电动车，作为代表中国的新兴产业，在世界范围内掀 起狂澜。 然而主流的MoE架构大模型，却苦于其结构上的"先天不足"：巨大的硬件成本与多重拖累效 率的环节，使得中国企业在这场芯片堆砌与效率挖掘的苦径上难以提速。 作为智能基础设施提供商，华为在这场战役中另辟蹊径，利用其在数学算法和工程领域的深 厚积累，为DeepSeek显著提升了效率及用户体验。 山就在那里，但中国企业找到了不一样的登顶之路。 近期，虎嗅将打造《华为技术披露集》系列内容，全面揭秘超大规模MoE模型推理部署技 术，通过一连串的技术报告，首次全面披露技术细节。 希望本系列内容能为业界起到参考价值，也希望更多人能与华为一起，共同打造长期持续的 开放协作生态环境，让昇腾生态在中国茁壮成长。 《华为技术披露集》系列 VOL.4 ：通信优化 01 大模型的推理， 就只是算力吗？ 大语言模型(Large Language Models, LLMs) 自从其问世以来，便迅速成为全球科技领域乃至 整个社会的焦点。 根据Scaling law，大语言模型的能力与其参数量的对数正相关，因此大语言模型的参数 ...

" MoE模型推理面临的3大通信难题，被通信尖子生华为逐一突 破，未来将进一步优化。 " 作者丨李希 大语言模型 (Large Language Models, LLMs) 自从其问世以来，便迅速成为全球科技领域乃至整个社会 的焦点。根据 Scaling law ，大语言模型的能力与其参数量的对数正相关，因此大语言模型的参数规模也 在指数级增长。随之而来的，是大语言模型部署形态的变化，从神经网络时代的单卡部署，到稠密模型时 代的多卡 / 单节点部署，再到以最近发布的 DeepSeek V3/R1 模型为代表的混合专家（ Mixture of Experts, MoE ）模型，它甚至会采用数百卡组成的集群和超节点来部署。 而在这基于集群的大模型推理中，集合通信操作就像是一群工人协作盖房子时传递材料和信息的方式，能 让多个计算节点高效配合完成任务。有一些常用集合通信操作，比如全量规约（A ll Reduce）可以想象 成一群工人各自收集了不同区域的建筑材料数据，全量规约就是把所有工人手里的数据汇总到一个地方， 进行求和、求平均值等计算。 大模型的推理，就只是算力吗？ 在大模型里，多个计算节点可能各自计算了一部分参 ...

机器之心发布 机器之心编辑部 在今年 2 月的 DeepSeek 开源周中，大模型推理过程中并行策略和通信效率的深度优化成为重点之一。 近日， 华为数学家出手，祭出 FlashComm，三箭齐发，解决大模型推理通算难题 ： 随着大语言模型（Large Language Models, LLMs）规模的指数级扩张，其部署形态也随之变化，显卡配置朝着规模化、集约化演进。从神经网络时代的单卡部署， 到稠密模型时代的多卡 / 单节点部署，再到以最近发布的 DeepSeek V3/R1 模型为代表的混合专家（Mixture of Experts, MoE）模型，大语言模型甚至会采用数百卡 组成的集群和超节点来部署。 可以说，模型推理早已不是「单兵作战」，而是一场高协同的「群体作战」。而在这基于集群的大模型推理中， 集合通信操作就像是一群工人协作盖房子时传递 材料和信息的方式，能让多个计算节点高效配合完成任务 。 由上可以看出， 集合通信操作是大模型推理中多个计算节点协作的「桥梁」，不同的并行策略（TP、DP、EP）通过这些操作实现高效的数据交互和计算，从而 加速大模型的推理过程 。 通信：Scaling law 头顶的 ...

Core Viewpoint - The company maintains stable operations despite a decline in revenue, with high growth rates in net profit attributable to shareholders and net profit excluding non-recurring items. The domestic computing power business continues to grow rapidly, and overseas capacity is being continuously improved, leading to an optimistic outlook for the company's future development, resulting in an upward adjustment of the target price and a maintained buy rating [1]. Financial Performance - The company reported a revenue of 28.549 billion yuan for 2024, a year-on-year decrease of 8.29%. The net profit attributable to shareholders was 703 million yuan, reflecting a year-on-year increase of 39.05%, while the net profit excluding non-recurring items was 653 million yuan, up 43.70% year-on-year. For Q1 2025, revenue was 4.049 billion yuan, down 19.97% year-on-year, with a net profit of 55.2 million yuan, an increase of 33.08% year-on-year, and a net profit excluding non-recurring items of 53.95 million yuan, up 28.96% year-on-year [2]. Profit Forecast - The company updated its profit forecast, expecting net profits attributable to shareholders to be 1.006 billion yuan, 1.167 billion yuan, and 1.344 billion yuan for 2025-2027, respectively. Corresponding EPS for 2025-2027 is projected to be 0.85 yuan, 0.99 yuan, and 1.13 yuan. Based on industry average valuation and prudent considerations, a PE of 35x for 2025 is given, raising the target price to 29.72 yuan from the previous 24.72 yuan, while maintaining a buy rating [2]. Business Segments - In the optical communication sector, the company achieved revenue of 22.313 billion yuan from communication system equipment, down 6.19% year-on-year, with a gross margin of 19.46%, up 0.26 percentage points. Revenue from fiber and cable was 4.699 billion yuan, down 12.77% year-on-year, with a gross margin of 25.35%, up 4.16 percentage points. Data network product revenue was 1.144 billion yuan, down 27.44% year-on-year, with a gross margin of 43.92%, up 2.56 percentage points. The company secured a position as a core supplier for the 400G network by winning bids from three major operators [3]. Computing Power Business - The domestic computing power revenue reached nearly 6 billion yuan, with a compound annual growth rate exceeding 100% over the past two years. The company’s Longjiang Computing has been certified as a leading partner for Huawei's Kunpeng and Ascend products, ranking in the top three in the ecosystem, indicating potential benefits from the growth in domestic computing power demand [3]. Global Expansion - The company has initiated trial production at its R&D and manufacturing bases in Thailand and Hungary, which is expected to help mitigate trade risks [3]. Cost Management - In 2024, the company's sales, management, and R&D expense ratios were 6.59%, 1.54%, and 10.09%, respectively, showing year-on-year changes of +0.85 percentage points, +0.44 percentage points, and -1.40 percentage points. The combined expense ratio decreased by 0.11 percentage points year-on-year, indicating effective cost control amid declining overall revenue [3]. Catalysts - The demand for domestic computing power is expected to exceed expectations, serving as a potential catalyst for growth [4].

机器之心发布 机器之心编辑部 在今年 2 月的 DeepSeek 开源周中，大模型推理过程中并行策略和通信效率的深度优化成为重点之一。 近日， 华为数学家出手，祭出 FlashComm，三箭齐发，解决大模型推理通算难题 ： 随着大语言模型（Large Language Models, LLMs）规模的指数级扩张，其部署形态也随之变化，显卡配置朝着规模化、集约化演进。从神经网络时代的单卡部署， 到稠密模型时代的多卡 / 单节点部署，再到以最近发布的 DeepSeek V3/R1 模型为代表的混合专家（Mixture of Experts, MoE）模型，大语言模型甚至会采用数百卡 组成的集群和超节点来部署。 可以说，模型推理早已不是「单兵作战」，而是一场高协同的「群体作战」。而在这基于集群的大模型推理中， 集合通信操作就像是一群工人协作盖房子时传递 材料和信息的方式，能让多个计算节点高效配合完成任务 。 有一些常用集合通信操作，比如 全量规约（AllReduce） 可以想象成一群工人各自收集了不同区域的建筑材料数据，全量规约就是把所有工人手里的数据汇总到 一个地方，进行求和、求平均值等计算。在大模型里，多个计算 ...

Core Viewpoint - The successful communication support for the recent Wuaizao Music Festival in Zaozhuang was achieved through effective planning and execution by Zaozhuang Unicom, ensuring seamless connectivity for attendees [1]. Group 1: Preparation and Deployment - Zaozhuang Unicom established a "Party Member Pioneer Special Guarantee Team" to integrate party work with communication support tasks, completing 12 rounds of venue inspections and identifying 3 weak signal points [3]. - The team utilized big data technology for traffic peak prediction and led a network upgrade project, enhancing 24 base station devices and adding 24 new carriers [3]. Group 2: Intelligent Scheduling - The core network parameters were optimized by 12 items, and an intelligent traffic scheduling system was deployed for real-time resource allocation [5]. - The venue had 3 existing 4G and 5G stations each, with 13 4G cells and 9 5G cells, and an additional 24 devices were installed on both sides of the venue [5][6]. Group 3: Performance and Results - After one week of optimization, the 5G download speed reached 232.92 Mb and upload speed 136.76 Mb, with zero interruptions in video streaming and calls [8]. - The event supported 23,000 attendees in sharing content in real-time, achieving over 1 million interactions on social media, and meeting the goal of "zero faults, zero complaints" [8].

hyzqdatemark 2025 年 05 月 22 日 政策发力，经济稳增向好 ——2025 年 4 月经济数据点评 投资要点： 证券分析师 廖志明 SAC：S1350524100002 liaozhiming@huayuanstock.com 证券研究报告 固收定期报告 请务必仔细阅读正文之后的评级说明和重要声明 4 月消费改善。4 月社会消费品零售总额 3.7 万亿元，同比增长 5.1%，环比增长 0.2%。 4 月升级类商品与绿色消费成为核心驱动力，金银珠宝、体育娱乐用品、家用电器、 家具分别同比+25.3%/+23.3%/+38.8%/+26.9%。1-4 月服务零售额同比增长 5.1%。 联系人 投资增速略微放缓。1-4 月全国固定资产投资（不含农户）14.7 万亿元，同比增长 4.0%，增速较 1-3 月份放缓 0.2pct。前四个月全国基础设施投资同比增长 5.8%， 制造业投资同比增速达 8.8%。房地产行业整体仍承压，但结构性改善信号显现，1-4 月新建商品房销售面积同比降幅较一季度收窄 0.2pct 至 2.8%。 4 月出口数据展现出较强韧性，在复杂国际环境中实现超预期增长。4 月在 ...

金融界2025年5月22日消息，国家知识产权局信息显示，天津七一二移动通信股份有限公司申请一项名 为"一种CIR报文结构翻译的方法"的专利，公开号CN120017214A，申请日期为2024年12月。 专利摘要显示，本发明公开了一种CIR报文结构翻译的方法，属于铁路通信技术领域。该方法通过接收 模块、循环字段赋值模块和报文接收类模块实现CIR报文的动态解析与翻译。接收模块用于接收原始报 文的Byte数组、偏移量、字段标题集合及字段大小集合等参数；循环字段赋值模块通过解析字段注解和 Byte数组，利用Java反射机制动态将解析结果赋值到接收类字段；报文接收类模块根据解析结果生成相 应的接收类对象并返回。该方法具有较高扩展性，可通过新增接收类适配新类型报文，无需针对每种报 文单独开发解析逻辑，从而提高开发效率，减少重复劳动。通过该方法，开发者可快速完成CIR报文翻 译与处理，提升铁路通信设备故障分析与监控的自动化水平。 天眼查资料显示，天津七一二移动通信股份有限公司，成立于2000年，位于天津市，是一家以从事计算 机、通信和其他电子设备制造业为主的企业。企业注册资本15200万人民币。通过天眼查大数据分析， 天津 ...

| 指数 | 收盘 | | 涨跌（％） | | | --- | --- | --- | --- | --- | | 名称 | （点） | 1 日 | 5 日 | 1 月 | | 上证指数 | 3388 | 0.21 | 2.92 | -0.48 | | 深证综指 | 2010 | 0 | 5.19 | -0.03 | | 风格指数 | 昨日表现 | 1 个月表现 | 6 个月表现 | | --- | --- | --- | --- | | | （％） | （％） | （％） | | 大盘指数 | 0.49 | 3.75 | -1.22 | | 中盘指数 | 0.19 | 1.92 | -6.84 | | 小盘指数 | -0.14 | 2.07 | -1.94 | 银行买债的底层逻辑、特征及展望——债市机构行为研究系列 之四 | 行业涨幅 | 昨日涨 | 1 个月涨 | 6 个月涨 | | --- | --- | --- | --- | | 贵金属 | 4.33 | -8.8 | 21.8 | | 商用车 | 3.57 | 5.07 | -1.67 | | 电池 | 2.99 | 12.34 | -5.55 | ...