一次算力政策研讨实录：算力调度的七个问题

Core Insights - The next five years (2026-2030) are considered a critical phase for the development of China's computing power industry, driven by the rapid growth of the AI sector which demands higher computing capabilities [1] - Key challenges facing the industry include limited capital expenditure and restricted supply of advanced AI chips, prompting companies to explore more efficient and cost-effective systems that integrate chips, algorithms, energy, and networks [1] Group 1: Policy and Industry Discussions - The National Information Center has organized discussions focusing on the value of "computing power scheduling" for industry development, involving experts from telecom operators, research institutions, and consulting firms [2] - Seven key issues were identified during these discussions, emphasizing the importance of efficient resource allocation in the computing power sector [2] Group 2: Understanding Computing Power Scheduling - "Computing power scheduling" is recognized as a method to optimize existing computing resources, reduce idle capacity, and achieve reasonable allocation, which is more complex than traditional resource scheduling like water or electricity [3][4] - The construction of a national integrated computing power network requires a comprehensive approach that combines government initiatives, standard-setting, and market operations to ensure efficient functioning [3] Group 3: Current State and Future Directions - As of the end of 2024, China's computing power centers are expected to exceed 9 million standard racks, with a computing power scale of 280 EFLOPS, ranking among the top globally [3] - The average energy utilization efficiency (PUE) of computing centers has improved to 1.46, indicating progress in resource efficiency [3] Group 4: Challenges and Considerations - There are concerns regarding the network costs associated with cross-regional scheduling, which may offset the benefits of lower electricity prices in the west [20] - Data security remains a significant issue, as clients worry about the privacy and sovereignty of their data when it is scheduled outside local environments [20] Group 5: Importance of Scheduling in the Integrated Network - Computing power scheduling is viewed as the "nervous system" of the national integrated computing power network, enhancing the flexibility of data centers and improving overall resource utilization [21][23] - Effective scheduling can lead to the integration of fragmented computing resources, reducing waste from redundant infrastructure [21] Group 6: Construction of a Unified Scheduling Network - Building a national computing power scheduling network involves more than just infrastructure; it requires standardization, platform development, and innovative mechanisms [24] - A successful network will necessitate a unified standard for computing power, a multi-level scheduling platform, and a sustainable operational service mechanism to facilitate rapid matching and flexible transactions [24][25]