机器之心发布 机器之心编辑部 现如今，随着参数规模的指数级增长，大语言模型（LLM）的能力边界不断被打破，AI 的智力正在经历快速跃迁。但随之而来的是，大模型在落地过程中面临着 一系列推理层面的难题，比如推不动、算不起、部署慢，导致推理成本高昂，性能冗余浪费严重。 因此，大模型推理的「速度」与「能效」成为所有算力厂商与算法团队绕不开的核心命题，如何让它们真正「跑得快、用得省」亟需全新的解法。这显然不仅仅 是工程挑战，更要在承接大模型推理压力的同时，在能效、延迟、成本等多方面实现可控与优化。 在这一背景下，华为团队和昨天一样（参考： 帮大模型提速 80%，华为拿出昇腾推理杀手锏 FlashComm，三招搞定通算瓶颈 ），用数学补物理，给出了一份深度 融合软硬件的系统性方案！ 他们基于昇腾算力，正式发布了三项重要的硬件亲和算子技术研究，带来了大模型推理速度与能效的双重革命 。具体包括如下： 可以看到，华为团队着力通过对大模型推理中关键算子的重构优化，实现能效、多卡协同和速度三大维度的全面突破。 作为 AI 大模型执行计算的「原子级工具」，算子如同乐高积木中的基础模块，负责从加减乘除到特征提取的一切核心操作。它们不 ...

Core Insights - PLP has commenced construction of a new multi-purpose facility in Wieprz, Poland, set to replace operations in Bielsko-Biała and enhance manufacturing capabilities by integrating modern engineering, operations, and sales support spaces, with completion expected in 2026 [1][2] - The new facility in Poland will serve as a key European hub for PLP's core product lines and services, reflecting the company's commitment to long-term growth in the European market [4] - PLP is also expanding its operations in Southern Europe by relocating to a larger facility in Seville, Spain, driven by rising demand and the need to scale production [2][3] Poland Facility Highlights - The new facility in Wieprz will feature a 30% increase in production space and a 50% increase in warehouse space, along with a world-class research and testing laboratory [7] - Modern offices and enhanced employee amenities will be part of the new work environment [7] Spain Facility Highlights - The Seville facility will see a 250% increase in operational space and a 240% increase in office capacity, allowing for team growth and collaboration [8] - Expanded manufacturing lines will support a broader product portfolio, and improved workspaces will enhance employee amenities [8] Strategic Vision - These investments are aligned with PLP's broader strategic vision to respond to the accelerating pace of global infrastructure projects, including grid modernization, renewable energy, and high-speed broadband [4]

" 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 ...

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 ...

证券代码：600498 证券简称：烽火通信 公告编号：2025-027 转债代码：110062 转债简称：烽火转债 烽火通信科技股份有限公司 关于参加中国信息通信科技集团有限公司 本公司董事会及全体董事保证本公告内容不存在任何虚假记载、误导性陈述或者重 大遗漏，并对其内容的真实性、准确性和完整性依法承担法律责任。 重要内容提示 ? 会议召开时间：2025 年 5 月 29 日（星期四）14:30-17:00 烽火通信科技股份有限公司（以下简称"公司"）已于 2025 年 4 月 26 日、 于广大投资者更全面、深入地了解公司 2024 年度、2025 年第一季度经营成果、 财务状况，根据公司间接控股股东中国信息通信科技集团有限公司（以下简称"信 科集团"）统一安排，公司将与信科集团下属的其他 5 家上市公司共同参加信科 集团 2024 年度暨 2025 年第一季度集体业绩说明会，通过视频和网络文字互动的 方式，与广大投资者进行互动交流。 董事、总裁：蓝海 副总裁、财务总监、董事会秘书：杨勇 一、说明会类型 本次集体业绩说明会由信科集团举办，旨在搭建一个坦诚、开放的沟通平台， 通过管理层与投资者的深度交流，全面 ...

Group 1: Company Initiatives - Ericsson has partnered with Zain Jordan to implement a Business Support Systems (BSS) transformation initiative aimed at enhancing digital services and customer experiences while increasing operational agility [1] - The initiative will transition Zain Jordan's existing BSS framework to a cloud-native architecture, aligning with the demands of the telecom and IT landscape [1][3] - The upgrade will expand Zain Jordan's current Ericsson Charging System, introducing new features hosted on Ericsson's Cloud Native Infrastructure Solution, enabling a catalog-based business model for improved customer service [2] Group 2: Operational Benefits - The transformation is expected to accelerate service delivery, reduce operational costs, and improve time to market, thereby enhancing operational flexibility and paving the way for 5G monetization [3] - This initiative supports Zain Jordan's broader digital transformation goals and contributes to national efforts to advance the digital economy [3] Group 3: Market Position and Financial Performance - Ericsson is focusing on 5G system development and has undertaken various initiatives to position itself for market leadership in this area, with innovative solutions reshaping connectivity across sectors [4] - The company is expected to benefit from an increasing customer base, which is likely to generate higher revenues in upcoming quarters, potentially leading to improved financial performance and stock price appreciation [5] - Over the past year, Ericsson's shares have gained 48.2%, outperforming the industry's growth of 40.1% [6]

金磊 发自 凹非寺 量子位 | 公众号 QbitAI 昨天的文章已经提到，昇腾超大规模MoE模型推理部署技术在本周会有持续的技术披露，果然第二天的技术报告又如期而至了。前情提要： 《华为 +DeepSeek，推理性能创新高！技术报告也公布出来了》 要问最近哪个模型最火， 混合专家模型 （MoE，Mixture of Experts）绝对是榜上提名的那一个。 它的巧妙之处，就在于把不同的任务分配给擅长处理的 专家网络 ，让整个系统性能得以提升。 但你知道吗？ 正是这个关键的专家网络，也是严重影响系统推理性能的因素之一。 因为在大量任务来临之际（尤其是超大规模时），MoE并不是以"雨露均沾"的方式去分配——专家网络们的 负载均衡问题 ，就会显得尤为 突出。 这个问题的根源，是因为某些专家网络总是被频繁调用（ 热专家 ），而另一些专家网络则鲜有机会派上用场（ 冷专家 ）。 没错，MoE里的"专家们"也是有冷热之分的，而且被调用频率的差距甚至可以达到 一个数量级以上！ 如此负载不均衡的现象，就会导致整个系统推理的时间被延长，以及还有资源利用率、系统性能受限等问题。 那么此局又该如何破解？ 别急， 华为团队 已经给出了 ...

Core Viewpoint - The emergence of the Harmony operating system is seen as a historical inevitability, driven by the need for a new technological ecosystem in response to the declining innovation vitality of the US-led single-core technology ecosystem [1][14][17]. Group 1: Development and Challenges of Harmony - Harmony OS is part of Huawei's "Root Technology Six Series," with the goal of making it widely recognized among consumers [2][3]. - The development of Harmony OS has faced delays, with the true version only being released in March and May of this year, indicating that the journey is just beginning [4][10]. - The system's growth is contingent on user adoption, with a target of reaching 100 million users primarily through the domestic market [32]. Group 2: Ecosystem and Market Dynamics - The ecosystem surrounding Harmony OS is crucial for its success, requiring a collaborative effort from various stakeholders, including developers and enterprises [27][31]. - The relationship between Harmony and WeChat reflects the complex dynamics of competition and cooperation within the ecosystem, highlighting the need for mutual survival [29][31]. - The development of Harmony is not just a Huawei initiative but a broader societal and industrial mobilization process [25]. Group 3: Global Context and Strategic Importance - The current technological landscape is characterized by a need for competition, as the single-core ecosystem has led to stagnation and monopolistic practices [15][14]. - The development of Harmony is positioned as a necessary step for China's high-tech self-reliance and a response to the geopolitical pressures faced by Huawei [20][14]. - The historical context of the internet's evolution is drawn parallel to Harmony's development, emphasizing the importance of a public goods approach to technology [9][12]. Group 4: Future Outlook and Vision - The vision for Harmony is to provide a viable alternative to existing systems, aiming to create a competitive landscape that fosters innovation [22][15]. - The expectation is that Harmony will eventually exceed the initial goal of capturing one-third of the market, although it currently faces the challenge of establishing a new ecosystem [23][24]. - The long-term success of Harmony will depend on its ability to adapt and grow organically, avoiding the pitfalls of rapid, unsustainable user influx [26][25].

Core Insights - AmpliTech Group, Inc. reported its Q1 FY2025 financial results, showing a revenue of $3.6 million, a 57% increase compared to Q1 FY2024, marking the strongest quarterly performance since Q4 FY2023 [6] - The company has provided optimistic revenue guidance for FY2025, expecting at least $21 million in revenue, more than double the previous fiscal year's sales, supported by a record backlog of $19.6 million and significant customer demand [3][4] Financial Performance - Revenue for Q1 FY2025 reached $3.6 million, reflecting a 57% increase over Q1 FY2024 [6] - The company ended the quarter with $19.1 million in cash and receivables, $24.6 million in working capital, and zero debt, indicating a strong financial position [6] - AmpliTech has booked $12 million in firm orders from Letters of Intent totaling over $118 million, with all booked revenue scheduled for delivery in FY2025 [3] Strategic Initiatives - The company achieved major operational milestones in the first four months of FY2025, including securing the largest backlog in its history and enhancing its intellectual property portfolio [4] - AmpliTech's certified ORAN 5G radios are being deployed globally, with shipments to a Tier 1 MNO in Canada already underway [4] - The core division expanded its proprietary LNB product line targeting the satellite market, with projections indicating LNB revenues may match core LNA sales within 12 months [7] Market Position and Growth Outlook - AmpliTech is positioned for successive quarters of growth, with a healthy balance sheet and an expanding portfolio of proprietary 5G and satellite technologies [8] - The company continues to build its global brand presence through participation in key industry events, enhancing its market visibility [7]