Workflow
国产化技术路径
icon
Search documents
高通量以太网联盟主席蔡德忠:破局AI算力瓶颈,以“慢功夫”换“真落地”
Huan Qiu Wang· 2025-08-25 02:14
【环球网科技报道 记者 林迪】近日,在第21届CCF全国高性能计算学术大会期间,由阿里云与中国科学院计算技术研究所联合发起的"高通量以太网 (ETH+)联盟"集中展示了在AI算力网络互联领域的重大突破,发布了涵盖协议标准、核心芯片、系统架构在内的一系列国产化成果,标志着该联盟在构 建自主可控、高性能、可扩展的智算网络基础设施方面迈出了关键一步。 算力跃迁的 " 阿喀琉斯之踵 " :网络互联瓶颈 随着AI大模型参数量的指数级增长,单一GPU已无法满足训练需求,须通过"Scale-Out"(横向扩展)和"Scale-Up"(纵向扩展)的方式,将成百上千个GPU 连接成一个超级计算集群。然而,这种并行计算模式带来了海量且密集的GPU间数据交换需求。相比传统通用计算,AI训练任务对网络带宽的要求通常高 出两个数量级。 更严峻的挑战在于,大模型训练中的数据同步具有明显的周期性。任何环节的性能短板——无论是网络链路拥塞还是设备故障——都可能成为整个集群 的"阿喀琉斯之踵",导致算力无法线性扩展,严重影响训练任务的进度与稳定性。业界普遍认为,如何构建一个能长期维持高带宽、低延迟和稳定性能的互 联体系,是确保集群算力随规模近 ...
2025制造行业(青岛)数智峰会举行
Qi Lu Wan Bao· 2025-05-17 06:34
Core Insights - The summit "Intelligent Manufacturing Cloud, Intelligent Computing Future" was held in Qingdao, focusing on the integration of industrial manufacturing with IDC computing power and AI models, highlighting the importance of digital transformation in the manufacturing sector [1][8] - The collaboration between Shandong Unicom and Beijing Parallel Technology aims to enhance industrial model training efficiency and reduce overall computing costs through deep integration of technology services and resource allocation [6] Group 1: Event Overview - The summit attracted over 200 attendees, including key figures from Shandong Unicom and Beijing Parallel Technology, emphasizing the significance of the event in promoting digital upgrades in manufacturing [1] - Discussions at the summit included topics such as domestic technology paths, general artificial intelligence development, and the future of intelligent manufacturing [8] Group 2: Shandong Unicom's Initiatives - Shandong Unicom is focusing on building computing network capabilities through its "YaoSuan" computing transaction scheduling platform and the China Unicom (Qingdao) Intelligent Computing Center, aiming to create an integrated AIDC service system [4] - The company plans to accelerate the construction of computing networks and develop a new information service system that combines computing power with capabilities to meet the digital economy's infrastructure needs in Shandong Province [4] Group 3: Beijing Parallel Technology's Role - Beijing Parallel Technology has 18 years of experience in the computing service field, and its partnership with Shandong Unicom is expected to enhance industrial model training efficiency [6] - The collaboration aims to lower comprehensive computing costs for enterprises, showcasing the potential benefits of combining technology services with resource allocation [6] Group 4: Key Discussions and Future Outlook - Experts at the summit discussed advanced topics such as industrial model capabilities, intelligent computing services, and the integration of supercomputing, showcasing real-world applications for intelligent manufacturing upgrades [8] - The successful hosting of the summit is seen as a catalyst for collaboration in AI and industrial manufacturing, contributing to the strategic goals of becoming a manufacturing and digital powerhouse in China [8]