Recently, the 17th ACM International Conference on Future and Sustainable Energy Systems (E-Energy 2026) was held in Banff, Alberta, Canada. The research project outcome paper (LLM-GreenServe) from the “Carbon Neutrality Practice Course” team at the Training Center was accepted by the AI-DEEDS Workshop at E-Energy 2026 for publication as a formal research paper.
The research paper was completed by the Carbon Neutrality Practice Course project team from the summer of 2025: Cui Zhihua (School of Environment), Li Hanglong (Department of Engineering Physics), Sun Shuoyang (Department of Industrial Engineering), and Ren Pinze (Tanwei College / Department of Chemical Engineering). The supervising instructors were Senior Engineer Chen Zhen, Associate Professor Lin Weiran, and Professor Li Shuangshou from the Training Center; Associate Research Fellow Li Chao from the National Research Center for Information Science and Technology; and Assistant Professor Wu Yinjun from the School of Computer Science, Peking University.
Figure 1. Course project team presenting the paper
Figure 2. The E-Energy AI-DEEDS Workshop in progress
The research outcome LLM-GreenServe was developed from the Carbon Neutrality Practice Project (research on energy consumption of AI computing facilities). This work proposes an LLM Green Serving system, which is based on an efficient power-aware PD ratio scheduling system and improves the power utilization efficiency of LLM serving systems.
LLM-GreenServe collects real-time data such as GPU power consumption and utilization, and integrates PID control to dynamically balance the allocation of Prefill and Decode tasks, thereby alleviating imbalances in hardware resource utilization. Empirical validation shows that, on the Qwen3-7B, 14B, and 32B series models, by using inputs from diverse tasks, it analyzes the impact of task types (such as long input/short output) on energy consumption optimization, providing a basis for scenario-based scheduling. Experiments verify the effectiveness of the strategy, with significant energy efficiency improvements for medium and larger models (14B/32B), saving up to 17.3%, especially in the non-chunked prefill mode.
This work explores the intersection of artificial intelligence and machine learning with green energy. In response to the rapid growth of energy consumption in LLM inference, it helps alleviate the significant tension between the expansion of large-model services and the national goals of “carbon peaking and carbon neutrality.” The proposed optimized GreenServe service scheduling represents a valuable exploration at the intersection of AI and machine learning and green energy.
E-Energy 2026 is supported by the Association for Computing Machinery (ACM) and has been successfully held for 16 previous editions, making it a long-established and influential conference in the fields of energy systems and information systems. The conference was held from June 22 to 25 in Banff, Canada. The AI-DEEDS Workshop is a workshop within E-Energy dedicated to AI-driven energy efficiency in dynamic systems.
