Sou Hu Cai Jing·2025-09-06 03:13Core Insights - The report titled "Computing Power Metropolitan Area Network White Paper (2025 Edition)" focuses on the explosive demand for computing power driven by generative AI and outlines the development path and technical directions for computing power metropolitan networks [1][9][10]. Group 1: Industry Development and Policy Trends - The rapid growth of computing power demand is attributed to advancements in AI, particularly large models like DeepSeek, which significantly reduce training and inference costs [9][17]. - By 2025, China's intelligent computing power is expected to reach 1037 EFLOPS, with a projected CAGR of 46.2% until 2028, when it may reach 2782 EFLOPS [1][17]. - The Chinese government has introduced various policies to optimize computing power infrastructure, including the "Digital China Construction 2025 Action Plan" and the "Computing Power Interconnection Action Plan" [17][18]. Group 2: Computing Power Metropolitan Network Demand - The computing power metropolitan network aims to provide efficient services for local users by integrating computing resources and ensuring high-speed, lossless data transmission [19][21]. - Key requirements for the network include handling massive data transfers (TB/PB level), supporting remote training with minimal performance degradation, and enabling cross-cluster collaborative training [22][23]. - The network must also facilitate cloud-edge collaboration and real-time inference delivery, ensuring high reliability and low latency [22][30]. Group 3: Network Architecture and Key Technologies - The architecture of the computing power metropolitan network includes components such as computing PODs, cloud network POPs, and export functional areas, utilizing technologies like SRv6 and EVPN [1][21]. - Key technologies identified include elastic bandwidth allocation (100 Mbps to 100 Gbps), ultra-high throughput capabilities, and wide-area lossless transmission [1][21]. - The network is designed to support high-capacity links (400G/800G) and ensure efficient resource utilization while maintaining high reliability and low latency [29][30]. Group 4: Typical Applications - The report outlines several application scenarios, including efficient data entry for massive datasets, remote training with data separation, and collaborative training across clusters [22][24]. - Specific use cases highlight the need for high-performance computing services in sectors such as government, finance, and healthcare, emphasizing the importance of data security and compliance [19][30]. - The computing power metropolitan network is positioned to accelerate digital transformation and support the rapid development of the digital economy [21][22].