国泰海通｜产业：华为盘古大模型与昇腾AI计算平台，共同构建软硬一体的AI技术体系

Core Viewpoint - Huawei is exploring a path to build its full-stack AI competitiveness through soft and hard collaborative innovation, transitioning from merely catching up with industry SOTA models to customizing model architectures to better leverage its self-developed Ascend hardware [1][2]. Group 1: AI Development Strategy - Huawei's AI development strategy has shifted towards a dual evolution path that addresses systemic issues in the large-scale application of AI models, focusing on a technology system composed of hardware-software collaborative architecture, operators, and software stacks [1]. - The evolution of the Pangu large model aims to solve efficiency challenges in large-scale distributed systems, particularly addressing the systemic bottleneck of expert load imbalance in the transition from dense architectures to mixture of experts (MoE) sparse architectures [1][2]. Group 2: Innovative Paths for Large Models - Huawei has launched two innovative paths at the large model level: Pangu Pro MoE, which introduces a grouped expert mixture (MoGE) architecture to tackle load imbalance, and Pangu Ultra MoE, which optimizes model architecture through system-level enhancements to better adapt to Ascend hardware [2]. - The physical foundation for this software-hardware collaborative innovation is the new generation AI infrastructure CloudMatrix, which features a unified bus (UB) network that reduces performance discrepancies in cross-node communication [2]. Group 3: Hardware and Software Synergy - The development of CloudMatrix not only provides a physical basis for software innovations like the Prefill-Decode-Caching (PDC) decoupled architecture but also enables high parallelism and low latency in software through large-scale expert parallelism (LEP) and operator-level optimizations like AIV-Direct [2].