Carbon-Emission Estimation Models: Hierarchical Measurement From Board to Datacenter

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Provides a novel hierarchical GNN approach for datacenter carbon accounting, bridging chip-level behavior and facility-wide emissions with high accuracy.

🏢実務担当者:Enables granular carbon attribution for datacenter operations, directly supporting ESG reporting and energy efficiency optimization.

🏛政策担当者:Demonstrates a scalable tool for regulating datacenter emissions, supporting digital industry carbon neutrality targets.

Data centers have become a core contributor to global digital carbon emissions, with their carbon footprint growing 19% annually alongside the expansion of AI and cloud services. Traditional carbon accounting methods are either trapped in macro-level rough calculation based on Power Usage Effectiveness (PUE) or limited to micro-level hardware power consumption measurement, failing to establish a traceable correlation between chip-level energy behavior and datacenter-wide carbon emissions. To address this gap, this study proposes a Hierarchical Coupling Carbon Emission Estimation Model (HCCEEM) that integrates physical modeling and graph neural network (GNN)-based statistical aggregation. The model constructs a four-level traceability chain spanning board (chip), server node, rack cluster, and campus datacenter, and introduces a real-time load adaptation module to capture dynamic workload impacts. Validated on a 14-month dataset from a heterogeneous cloud datacenter, HCCEEM achieves an estimation accuracy of 95.7%, reducing mean absolute error (MAE) by 27.1% and 19.3% compared to PUE-based models and single-level machine learning models respectively. Moreover, the model realizes fine-grained attribution of carbon contributions across levels, revealing that chip-level dynamic power consumption drives 65.2% of server emissions, and rack-level cooling losses account for 33.8% of datacenter emissions. This research provides an interpretable, scalable tool for targeted carbon reduction, bridging the gap between hardware-level optimization and datacenter-wide carbon management. Specifically, HCCEEM exhibits remarkable applicability in high-load scenarios such as large language model (LLM) training and inference, where it can reduce carbon accounting errors by over 30% compared to conventional methods. For small and medium-sized datacenters with limited monitoring resources, the model’s modular design allows lightweight deployment by simplifying partial hierarchical modules without significant accuracy loss. Additionally, the hierarchical contribution quantification function of HCCEEM can directly support enterprises’ carbon disclosure and compliance reporting, aligning with the carbon neutrality requirements of the digital industry in various regions.

• semanticscholar https://doi.org/10.70393/6a69656173.333931first seen 2026-06-29 08:36:33