华为发布CloudMatrix384超节点，革新算力集群

Investment Rating - The investment rating for the communication equipment industry is "Positive" and maintained [8]. Core Insights - Huawei Cloud launched the revolutionary CloudMatrix 384 super node, which breaks through computing bottlenecks with a "poolable, equalized, and combinable" architecture, achieving dual leadership in interconnect bandwidth and memory bandwidth [2][11]. - The CloudMatrix 384 super node is expected to alleviate computing anxiety, with plans for tens of thousands of units to be deployed in the first half of the year, aiming to build a large cluster of over ten thousand nodes [11]. - The CloudMatrix 384 achieves a computing scale of 300 PFlops, a 67% increase compared to NVIDIA's NVL72, which has a computing scale of 180 PFlops [11]. - The service DeepSeek-R1, based on the CloudMatrix 384, can achieve a single-card decode throughput of over 1920 Tokens/s while maintaining a single-user transaction per second (TPS) level of 20 [2][11]. - Huawei Cloud's ecosystem has surpassed 12 million developers, 45,000 partners, and 12,000 cloud store applications, forming a comprehensive digital ecosystem across the entire industry chain [11]. Summary by Sections Event Description - On April 10, Huawei Cloud held an ecological conference in Wuhu, Anhui, where it launched the CloudMatrix 384 super node and announced its large-scale deployment in the Wuhu data center [5]. Event Commentary - The CloudMatrix 384 super node redefines AI infrastructure, transitioning from server-level to matrix-level resource supply models, significantly enhancing interconnect bandwidth and achieving a breakthrough in memory bandwidth [11]. - The architecture of the CloudMatrix 384 is designed to support the rapid deployment of industry applications, with over 160 third-party large models already adapted to its services [11].