华为云黄瑾：六大科技创新，昇腾云定义下一代AI基础设施

Core Viewpoint - The emergence of super-node architecture is a breakthrough in technology and engineering innovation, addressing the explosive growth in computational power demand for AI training and inference, as stated by Huawei Cloud's Vice President Huang Jin [1][3]. Group 1: Technological Innovations - Huawei Cloud's CloudMatrix 384 super-node features six key technological advantages: MoE affinity, network-enhanced computing, memory-enhanced computing, long-term reliability, peak-shifting resource usage, and on-demand availability [1][4]. - The evolution of AI model training has transitioned from small models on single cards to large models on AI clusters, with computational power demand increasing by 10,000 times [3]. - Over the past eight years, single-card hardware performance has increased by 40 times, while inter-node network bandwidth has only increased by 4 times, highlighting the communication bottleneck in AI model training [3]. Group 2: Implementation and Applications - The CloudMatrix super-node clusters have been deployed in Wuhu, Gui'an, and Inner Mongolia, showcasing the practical application of the new architecture [4]. - Huawei Cloud aims to transform AI technology value into practical applications across various industries, emphasizing the importance of AI as a universal technology [5]. - The company has adapted its Ascend AI cloud services to over 160 large models, assisting clients in model development, training, hosting, and application across sectors such as government, finance, retail, internet, transportation, and manufacturing [5].