训练效率不足 ，甚至 一半以上训练时间都浪费在"等待"上 。 现在，为了突破MoE的训练瓶颈，华为出手了： 构建了一套名为 Adaptive Pipe & EDPB 的优化方案，开启"上帝视角"，让MoE面临"交通拥堵"的训练集群， 实现无等待流畅运行。 MoE大规模训练难题：一半以上的训练时间在等待？ 实践已经表明，MoE模型训练集群的效率面临两方面挑战： 首先，是 专家并行引入了计算和通信等待 。 允中 发自 凹非寺 量子位 | 公众号 QbitAI Scaling Law之下，MoE（混合专家）如今已经成为各大模型厂商扩展模型能力的制胜法宝。 不过，在高效实现模型参数规模化的同时，MoE的 训练难题 也日益凸显： 当模型规模较大时，需要切分专家到不同设备形成并行（EP），这就引入额外All-to-All通信。 与此同时，MoE层绝大部分EP通信与计算存在时序依赖关系，一般的串行执行模式会导致大量计算单元空闲， 等待通信。 其次， 负载不均会引入计算和计算等待 。 MoE算法核心是"有能者居之"，在训练过程中会出现部分热专家被频繁调用，而冷专家使用率较低的情况。 同时，真实训练数据的长度不一，不同的模型层 ...

5月31日，在2025未来汽车先行者大会上，华为常务董事、终端BG董事长余承东表示，鸿蒙智行可能不 会再有第六"界"。他称："因为我们的能力做两三个界都已经很不容易了，做五个界非常非常难。一个 拳头还真的需要五个手指头，握上才有力量。所以也不可能有六个手指头，我们就到此暂时握紧了，一 起努力干，一起打天下。"(第一财经) ...

Core Viewpoint - Huawei has achieved significant advancements in training large models through its "Ascend + Pangu Ultra MoE" system, demonstrating a fully domestic and GPU-free training process that enhances computational efficiency and model performance [3][4][38]. Group 1: Technical Innovations - Huawei's training system has achieved a model training efficiency with a utilization rate (MFU) of 41% during the pre-training phase using the Ascend Atlas 800T A2 cluster [4][38]. - The Pangu Ultra MoE model consists of 718 billion parameters, featuring a unique architecture with 61 layers, including 58 MoE layers, and is designed for high performance and scalability [38][39]. - The system supports a high throughput of 35K Tokens/s during the reinforcement learning (RL) post-training phase, showcasing its capability to process complex tasks rapidly [39]. Group 2: Challenges Addressed - The report identifies six key challenges in the current MoE pre-training and RL post-training processes, including difficulties in parallel strategy configuration, communication bottlenecks, and uneven system load distribution [7][10][12][13]. - Huawei has developed a comprehensive end-to-end solution to address these challenges, focusing on optimizing training cluster utilization and enhancing communication efficiency [14][16][25]. Group 3: Specific Solutions - The first strategy involves improving training cluster utilization through intelligent parallel strategy selection and global dynamic load balancing, significantly enhancing overall training efficiency [16][23]. - The second strategy focuses on releasing computational power at the single-node level by optimizing training operators and enhancing memory management, achieving a twofold increase in micro-batch size [26][30]. - The third strategy introduces high-performance scalable RL post-training technologies, allowing for flexible deployment modes and doubling the utilization rate of RL post-training clusters [33][34].

Group 1 - Dongfeng Motor Group Co., Ltd. and Huawei Technologies Co., Ltd. have signed a comprehensive strategic cooperation agreement in Wuhan [1] - The collaboration will focus on areas such as automotive intelligence, enterprise digitalization, and intelligent upgrades [1] - The partnership aims to promote ecological co-construction and deepen cooperation across various fields [1]

Core Viewpoint - Shenzhen Zhenyou Technology Co., Ltd. has been awarded the second prize of the Shenzhen Science and Technology Progress Award for its project on "Key Technology Innovation and Industry Application of Hierarchical Heterogeneous Communication Integration and All-Domain Command and Dispatch" [1][2] Group 1: Project Overview - The awarded project focuses on the integration of various communication technologies, including wireless communication, satellite communication, and IoT, to create a comprehensive emergency management information system [1] - The project aims to enhance real-time perception of natural disasters and safety production accidents through a multi-dimensional sensing system, improving the timeliness, accuracy, and effectiveness of information acquisition [1] Group 2: Innovation and Impact - The project addresses bottlenecks in command and dispatch systems related to information sharing, emergency response, and task collaboration by utilizing big data and artificial intelligence [2] - The award is a recognition of the company's technological innovation capabilities, which is expected to enhance its core technology advantages and market competitiveness [2]

Group 1 - The appointment of the new U.S. ambassador to China, Qin Gang, reflects the deep contradictions in the Trump administration's China policy, as evidenced by Trump's simultaneous desire to visit China while imposing domestic pressures on Walmart and the Federal Reserve [1][3] - The U.S. federal debt has surpassed $36 trillion, and the 10-year Treasury yield has risen to 4.5%, causing anxiety in the White House, especially as China has been reducing its holdings of U.S. Treasuries [1][3] - Trump's pressure on Walmart and the Federal Reserve reveals his true policy intentions, as the Consumer Price Index has risen by 6.2% year-on-year due to tariffs on Chinese goods, ultimately burdening American consumers [3][5] Group 2 - The new ambassador, Qin Gang, has a close relationship with Trump and previously advocated for "supply chain decoupling," indicating a hawkish stance despite claims of strategic engagement [5][6] - The ongoing "truce" period in U.S.-China relations allows for potential negotiations, but China insists on equality in discussions, highlighting the failure of Trump's previous tariff strategies [5][6] - The shifting attitudes of the Trump administration are accelerating the "de-dollarization" process, as countries like Saudi Arabia and the UAE seek alternatives to U.S. influence [6][8]

Group 1 - Dongfeng Motor Group and Huawei have signed a strategic cooperation agreement to deepen their collaboration [1] - The partnership will focus on areas such as automotive intelligence, enterprise digitalization, and intelligent upgrades [1] - Both companies aim to build an ecosystem through this cooperation [1]

Core Viewpoint - The collaboration between Shenzhen Press Group and Huawei aims to establish an AI foundation for the media industry, facilitating intelligent transformation and exploring the "Shenzhen Press+" technology development model [1][3]. Group 1: Collaboration Details - The partnership is a significant step for Shenzhen Press Group in advancing systemic transformation, aligning with the city's status and mission as a "window of reform and opening up" [3]. - The agreement focuses on three key areas: restructuring the technological foundation, reengineering production processes, and upgrading international communication [3]. - The collaboration will enhance AI applications, data empowerment, talent cultivation, and domestic innovation, promoting deep integration of "Artificial Intelligence+" [3][5]. Group 2: Technological Innovation - Shenzhen Press Group plans to establish a "smart brain" for technological innovation, increasing strategic investment in technology and collaborating with leading tech companies like Huawei [5]. - A new Technology Innovation Department will be set up at the group headquarters to oversee overall technology development planning [5]. - The group aims to create a three-tier transformation system of "strategy-technology-scenario," applying AI, big data, and blockchain across all content production and dissemination processes [5]. Group 3: Future Outlook - The partnership is committed to mutual benefits and exploring the deep application of AI technology in the media sector, driving the intelligent transformation of the industry [5].

"人们一度忘了这个保护区，忘了美洲豹曾是世代祖先口中的猛兽。"玛雅社群领袖本哈明·坎波斯告诉 新华社记者。在他的家乡墨西哥东南部尤卡坦的齐拉姆自然保护区，一场由数字科技带动的生态保护革 新正在悄然上演。 在5月22日国际生物多样性日前夕，记者一行凌晨5时从墨西哥尤卡坦州州府梅里达出发，驱车前往齐拉 姆自然保护区。经过两个多小时颠簸后，车队前方道路彻底隐没在灌木丛中，手机信号中断。记者一行 背起行囊，徒步进入保护区腹地，寻找坎波斯口中的美洲豹踪迹。 坎波斯生活的尤卡坦半岛不仅是玛雅文明发源地之一，也以丰富的生物多样性著称。其中，齐拉姆自然 保护区占地超过690平方千米，包括约525平方千米陆地和165平方千米海洋。保护区内有超过590种动植 物，其中有数十种濒危及重点保护物种。 新华社墨西哥梅里达5月21日电 行走拉美手记丨中墨科技合作为美洲豹打造"云端栖息地" 新华社记者翟淑睿 吴昊 据尤卡坦政府数据，截至2025年4月，上述项目已识别出147个物种，其中在当地首次记录并共确认16只 美洲豹活动踪迹。 跟着坎波斯的脚步往林中更深处去，土地越发柔软潮湿。这里的树干高大挺拔，一丛丛气生根粗壮繁 茂，枝叶在头顶交错 ...

Core Insights - The world's first 5G-A open-pit mine has been launched in Inner Mongolia, featuring a fleet of 100 unmanned electric mining trucks known as "Huaneng Ruichi" [1] - This project represents a significant advancement in achieving "vehicle-cloud-network" collaborative operations on a large scale [1] - The project is a collaboration among China Huaneng, XCMG Group, Huawei, and State Grid Intelligent Vehicle Networking Co., Ltd [1] Technology Implementation - The 5G-A network deployed in the Yimin mining area is the first of its kind for open-pit mining globally [1] - The network supports an uplink speed of 500 Mbps and a low latency of 20 milliseconds [1] - It enables over 300 unmanned mining trucks to operate continuously for 24 hours [1]