Core Viewpoint - Huawei is building a full-stack AI competitive advantage through software and hardware collaboration, leading to a technological revolution in the communication equipment industry [1] Group 1: Huawei's AI Strategy - Huawei's AI strategy has shifted from benchmarking SOTA models to customizing architectures for Ascend hardware, introducing two innovative pathways: Pangu Pro MoE and Pangu Ultra MoE [1] - These pathways address load imbalance issues and enhance hardware efficiency through a mixture of expert groups (MoGE) architecture and system-level optimization [1] Group 2: New AI Infrastructure - The new generation AI infrastructure, CloudMatrix, utilizes a unified bus (UB) network to create a distributed high-speed memory pool, reducing cross-node communication discrepancies [1] - It supports PDC separation architecture and large-scale expert parallelism (LEP), focusing on distributed system efficiency challenges as large models transition from dense to sparse MoE architectures [1] Group 3: Industry Implications - The communication equipment industry is evolving towards a fully collaborative technical system, with Huawei expanding its software and hardware innovation into AI system engineering [1] - The communication ETF (515880) tracks the communication equipment index (931160), which focuses on the manufacturing and related services of communication equipment, reflecting the overall performance of listed companies in this sector [1] - The index is characterized by high technical content and growth potential, making it a relevant investment focus for those interested in the communication equipment sector [1]

Group 1 - Huawei is building a full-stack AI competitiveness through soft and hard collaboration from large model design to infrastructure, shifting its AI development strategy from benchmarking industry SOTA models to self-developed Ascend hardware tailored model architecture [1] - Huawei has introduced two innovative paths at the large model level: Pangu Pro MoE, which addresses load imbalance through a mixture of experts (MoGE) architecture, and Pangu Ultra MoE, which achieves collaborative optimization of training and inference through system-level optimization for Ascend hardware [1] - The new generation AI infrastructure, CloudMatrix, features a unified bus (UB) network as its core technology, reducing cross-node communication performance discrepancies through a distributed high-speed memory pool, providing a physical basis for upper-layer software innovation [1] Group 2 - The communication ETF (515880) tracks the communication equipment index (931160), which mainly covers listed companies engaged in communication network infrastructure and terminal equipment, characterized by high technical content and R&D investment [1] - The industry allocation focuses on 5G, Internet of Things, and related fields to reflect the overall performance of listed companies in the communication equipment sector [1]

Group 1 - Huawei is building a full-stack AI competitiveness through soft and hard collaboration, shifting its strategy from benchmarking industry SOTA models to customizing model architecture for self-developed Ascend hardware [1] - Huawei has introduced two innovative paths at the large model level: Pangu Pro MoE and Pangu Ultra MoE, addressing load imbalance issues through the mixture of experts (MoGE) architecture and system-level optimization [1] - The new AI infrastructure CloudMatrix creates a distributed high-speed memory pool via a unified bus network, reducing performance discrepancies in cross-node communication, which provides a physical basis for upper-layer software innovation [1] Group 2 - The Growth Enterprise Market Artificial Intelligence ETF from Guotai (159388) tracks the Growth Enterprise Market Artificial Intelligence Index (970070), with a daily fluctuation limit of up to 20% [2] - The index selects listed companies involved in AI technology development and intelligent services from the Growth Enterprise Market, reflecting the overall performance of AI-related listed companies [2] - The index components cover various subfields, including software and hardware research and development, and intelligent application solutions, showcasing significant technological innovation attributes [2]

Group 1 - Huawei is building a full-stack AI competitiveness through soft and hard collaboration, transitioning from industry SOTA models to self-developed Ascend hardware tailored model architectures [1] - The Pangu Pro MoE adopts a mixture of experts (MoGE) architecture to address load imbalance issues, while Pangu Ultra MoE optimizes system-level adaptation for Ascend hardware [1] - The new AI infrastructure CloudMatrix constructs a distributed high-speed memory pool via a unified bus (UB) network, reducing cross-node communication discrepancies and supporting software innovations like PDC separation architecture [1] Group 2 - The software ETF (515230) tracks the software index (H30202), which selects listed company securities involved in software development, system integration, and internet services to reflect the overall performance of the software industry [1] - The index components cover application software, system software, and other segments within the information technology field, showcasing the technological innovation capability and market growth potential of software service companies [1] - Investors without stock accounts can consider the Guotai Zhongzheng All-Index Software ETF Connect A (012636) and Guotai Zhongzheng All-Index Software ETF Connect C (012637) [1]

Investment Rating - The report rates the computer industry as "Outperforming the Market" [32] Core Insights - Nvidia reported strong Q1 earnings with revenue of $44.1 billion, a 12% quarter-over-quarter increase and a 69% year-over-year increase, despite being affected by export controls [11][12] - The DeepSeek R1 model has completed a minor version upgrade, achieving top performance among domestic models and nearing international leaders [13][14] - Huawei's Pangu team launched the Pangu Ultra MoE model, addressing stability issues in training large-scale models, which signifies a successful practice of autonomous training using domestic computing power [15][16] Company Dynamics - Zhongke Chuangda announced a special loan commitment of up to 70 million yuan for stock repurchase [3] - Kingsoft Office disclosed the results of its restricted stock incentive plan, with a total of 505,289 shares newly added, bringing the total share capital to 463,179,293 shares [23] - The report highlights the importance of companies in the Huawei supply chain and EDA software sector, suggesting a focus on firms like Softcom Power, Tuo Wei Information, and others [4]

Core Viewpoint - Huawei's advancements in MoE (Mixture of Experts) training systems demonstrate its leading capabilities in AI foundational technology and engineering implementation [1][2]. Group 1: MoE Training System Enhancements - Huawei has introduced new solutions for MoE training operators and memory optimization, achieving a 20% increase in system throughput and a 70% reduction in memory usage [2][7]. - The MoE framework is becoming a preferred path for tech giants aiming for more powerful AI systems [3]. - The unique architecture of MoE is key to overcoming computational bottlenecks in large-scale model training [4]. Group 2: Challenges in MoE Training - MoE model training faces significant challenges, particularly in single-node efficiency, due to low operator computation efficiency and memory constraints [10][11]. - The complexity of the expert routing mechanism leads to frequent operator dispatch interruptions, creating a Host-Bound bottleneck [12]. - The need for extensive model parameters results in high memory demands, often leading to out-of-memory (OOM) issues during training [13][15]. Group 3: Solutions and Innovations - Huawei has developed a comprehensive solution to address the challenges in MoE training, focusing on enhancing operator computation efficiency and memory utilization [17]. - The collaboration between Ascend and Kunpeng architectures has significantly improved training operator efficiency and memory usage [6][34]. - The implementation of three optimization strategies—"Slimming," "Balancing," and "Transporting"—has led to a 15% increase in overall training throughput for the Pangu Ultra MoE 718B model [20][21]. Group 4: Specific Operator Optimizations - FlashAttention optimization has improved performance by 50% for forward and 30% for backward processes through efficient computation order and reduced redundancy [23][25]. - Matrix multiplication operator enhancements have increased core utilization by 10% through optimized data transport strategies [26][28]. - Vector operator optimizations have resulted in performance improvements exceeding three times by minimizing data transport during reordering operations [30][32]. Group 5: Memory Optimization Techniques - The Selective R/S memory optimization technique has enabled a 70% reduction in activation memory during training by implementing fine-grained recomputation and adaptive memory management [46][49]. - The self-adaptive memory optimization mechanism focuses on maximizing the efficiency of memory usage relative to additional computation time [55][56]. Group 6: Industry Implications - Huawei's deep collaboration between Ascend and Kunpeng, along with its innovative operator acceleration and memory optimization techniques, provides an efficient and cost-effective solution for MoE training [58]. - These advancements not only eliminate barriers for large-scale MoE model training but also offer valuable reference paths for the industry [59].

Core Viewpoint - Huawei defines the benchmark for domestic large model training through technological innovation, achieving breakthroughs in computing power utilization and post-training throughput [1][4]. Group 1: Technological Innovations - Huawei's "Ascend + Pangu Ultra MoE" combination has unlocked a fully controllable training loop for domestic computing power and models, achieving industry-leading performance in cluster training systems [4][5]. - The pre-training phase saw the Ascend Atlas 800T A2 cluster's model training utilization (MFU) increase to 41%, while the post-training phase achieved a throughput of 35K Tokens/s on a single CloudMatrix 384 super node [5][36]. - Huawei disclosed key technologies in its technical report, highlighting the efficient integration of sparse MoE reinforcement learning post-training frameworks [6][7]. Group 2: Challenges in Current Training Processes - Six main challenges were identified in the current MoE pre-training and reinforcement learning post-training processes, including difficulties in parallel strategy configuration, communication bottlenecks, uneven system load distribution, excessive operator scheduling overhead, complex training process management, and limitations in large-scale expansion [10][11]. Group 3: Solutions to Enhance Training Efficiency - Huawei proposed a complete end-to-end solution to address these challenges, focusing on enhancing training cluster utilization through intelligent parallel strategy selection, deep integration of computation and communication, and global dynamic load balancing [12][14]. - The first strategy involved optimizing parallel configurations, achieving a deployment that included 16 pipeline parallelism, 8 tensor parallelism, and 32 expert parallelism [15][16]. - The second strategy focused on releasing computing power at the single-node level, doubling the micro-batch size (MBS) and optimizing operator scheduling to fully utilize Ascend node capabilities [20][21]. Group 4: Reinforcement Learning Innovations - Huawei introduced the RL Fusion training and inference co-card technology, which supports flexible deployment modes and achieves a doubling of cluster utilization in post-training [28][29]. - The design of a semi-asynchronous mechanism, StaleSync, allows different tasks to execute in parallel while maintaining model accuracy, resulting in a 50% increase in overall training throughput [30]. Group 5: Performance Metrics and Future Prospects - The Pangu Ultra MoE model, with 718 billion parameters, demonstrated high performance during training, achieving a model utilization rate of 41% and a throughput of 35K Tokens/s in post-training [35][36]. - The system is designed to support ultra-large-scale clusters and models, with expectations for future iterations to achieve even higher utilization rates [35][36].