并行科技赵鸿冰：如何最大化发挥算力效益？丨GAIR 2025

Core Viewpoint - The article emphasizes the rapid growth and evolution of the computing power market, highlighting the importance of building a robust computing service system from the user's perspective and the need for efficient resource integration and scheduling through computing networks [3][4][18]. Group 1: Computing Power Market Overview - The computing power market has experienced explosive growth across multiple scenarios and business types, evolving from supercomputing to intelligent computing forms, and from computing power leasing to computing networks [3][4]. - The current computing power market is characterized by four core business types: computing power leasing, computing power services, computing power operations, and computing power networks [3][22]. Group 2: User-Centric Computing Services - The company has developed a "factory-network integration" model, combining heavy asset investments in computing clusters with light asset expansion to connect various computing centers across the country [4][27]. - The computing power network can schedule over 2 million CPU cores and more than 50,000 GPU cards, serving over 160,000 users, with commercial output exceeding 20 billion core hours and nearly 200 million card hours [4][27]. Group 3: Key Trends and Future Directions - The compound annual growth rate of computing power is projected to reach 52.3%, driven by significant capital investments in AI infrastructure, despite concerns about a potential "computing power bubble" [5][6]. - The demand for inference capabilities is expected to drive the next wave of growth in the computing power market, with major clients already entering the stage of implementing inference business [11][12]. Group 4: Technological Innovations and Competitive Edge - The company has established a mature standard system for integrating computing resources, allowing for rapid access and networked output of computing power [7][8]. - A performance prediction model has been developed, achieving prediction errors of less than 2% in small-scale scenarios and single-digit errors in medium to large-scale scenarios, supporting user resource selection decisions [35][36]. Group 5: Market Position and Client Base - The company serves a diverse client base, including top universities and research institutions, with significant partnerships established to provide computing support for AI research [43][45]. - The computing power index is likened to essential utilities like water and electricity, indicating its foundational role in the digital economy, with a 1% increase in computing power expected to boost GDP by hundreds of billions [45].