Workflow
华为CloudMatrix384超节点发布，国产大规模算力集群首秀
东北证券·2025-04-17 07:16

Investment Rating - The report rates the industry as "Outperforming the Market" [5] Core Insights - Huawei's CloudMatrix 384 super node has been officially launched, marking the commercial deployment of a large-scale domestic computing cluster [1] - The CloudMatrix 384 super node achieves a total computing power that is 67% higher than NVL72, with network interconnect bandwidth increased by 107% and memory bandwidth by 113% [2] - The architecture of CloudMatrix 384 allows for a linearity of over 95% with thousands of cards, demonstrating significant breakthroughs in computing efficiency and engineering reliability [2] - The system-level thinking of Huawei is highlighted, focusing on enhancing chip performance through stacking and splicing rather than just single-point performance [3] - The UB-Mesh architecture proposed by Huawei aims to improve scalability, performance, cost-effectiveness, and availability in AI data centers [4] - The CloudMatrix 384 super node is positioned to compete with NVIDIA's NVL72, while the Ascend 910C single card is compared to the NVIDIA H100, indicating a breakthrough for domestic AI GPUs in pure training scenarios [4] Summary by Sections Product Launch - Huawei CloudMatrix 384 super node was launched at the Huawei Cloud Ecosystem Conference on April 10, 2025, and is operational at the Wuhu data center [1] Performance Metrics - Each card in the CloudMatrix 384 has a computing power of approximately 781.25 Tflops and a total memory bandwidth of 3200 GB/s, with interconnect bandwidth of 350 GB/s in a fully interconnected state [3] Network Architecture - The UB-Mesh architecture allows for a maximum of 1024 interconnected cards within a Pod, utilizing a 2D full interconnect topology to enhance performance and reduce reliance on expensive high-bandwidth switches [4] Market Positioning - The report suggests that Huawei's AI GPU products will continue to evolve, addressing the computing power anxiety in the industry and emphasizing the importance of the Ascend industrial chain [4]