| Author: | Zhang, Xiaoyang |
| Title: | Towards carbon-intelligent computing : carbon forecasting and accounting models for computer and large AI systems |
| Advisors: | Wang, Dan (COMP) |
| Degree: | Ph.D. |
| Year: | 2026 |
| Department: | Department of Computing |
| Pages: | xvii, 118 pages : color illustrations |
| Language: | English |
| Abstract: | The rapid expansion of artificial intelligence (AI) and large-scale computing systems has led to a significant increase in their carbon footprint, which includes both operational carbon (from electricity consumption) and embodied carbon (from hardware manufacturing). To enable the carbon reduction of carbon-aware systems, three critical challenges remain understudied: (1) accurate carbon intensity forecasting of electricity in complex cross-border power grids, (2) generalizable carbon intensity forecasting in carbon data-scarce regions, and (3) granular embodied carbon accounting of computer systems that captures spatial-temporal dynamics and carbon attribution methods. First, existing carbon intensity forecasting models are designed for regional or one-hop neighboring grids. They fail to capture the complex spatial and temporal dependencies in cross-border power grids, where electricity exchanges involve multiple countries and dynamic carbon flows. This leads to inaccurate forecasts, which in turn limit the effectiveness of carbon-aware scheduling in distributed computing systems. To solve this problem, we propose cross-border carbon flows, and such flows form a carbon network. The challenge is to learn the complex spatial and temporal dependencies that are involved. We propose a new CFCG model based on Graph Neural Networks (GNN). Our model has three features: (1) the carbon intensity variation shows periodic patterns in multiple levels of granularity, e.g., hourly, daily, and weekly. To capture the patterns across various granularities, we develop a multi-periodic pattern encoding scheme to transform the original carbon intensity sequence data into data with different periodic granularity settings (e.g., hourly, daily, weekly); (2) Cross-border power grids have carbon flows with spatial dependencies. Thus, in our CFCG model, we develop and integrate a GNN-based submodel (i.e., GNN layers) to learn the spatial dependencies; (3) Carbon flows have time dependencies. Thus, in our CFCG model, we develop and integrate an LSTM-based submodel (i.e., LSTM layers) to uncover the temporal dependencies. We further evaluate our model through two case studies where we use the model to support Large Language Model training and inferences to reduce their carbon footprint and show the potential carbon footprint reduction of up to 90.89%. Second, existing carbon intensity forecasting methods focus on data-rich regions, yet most areas lack sufficient carbon data due to carbon intensity's inherent measurement limitations (e.g., reliance on indirect estimation rather than direct sensor metering) and dependence on upstream entities like grid operators for data disclosure. We propose DGCFM, a Dual Graph empowered Carbon-domain Foundation Model, enabling cross-regional general carbon intensity forecasting, especially for data-scarce regions. DGCFM integrates a pre-trained Time Series Foundation Model (TSFM) to capture temporal patterns under data constraints, empowered by metadata-driven carbon hypergraph fine-tuning and a heterogeneous spatiotemporal graph to capture spatial dependency in carbon networks. Evaluated on 102 real-world datasets, DGCFM achieves 20.04% average accuracy improvement in low-data scenarios. Third, the embodied carbon of computer systems is the carbon emissions in the manufacturing process of hardware, which dominate the overall carbon footprint. Existing studies derive the embodied carbon through life cycle analysis (LCA) reports. Current LCA reports only provide the carbon emission of a product class, e.g. 28nm CPU, whereas a product instance can be made in various regions and time periods. Carbon emissions depend on the electricity generation process, which has spatial-temporal dynamics. Therefore, the embodied carbon of a product instance can differ from its product class. Additionally, different carbon attribution methods (e.g., location-based and market-based) can affect the carbon emissions of electricity, thus further affecting the embodied carbon of products. In this paper, we present new Spatial-Temporal Embodied Carbon (STEC) accounting models with dual attribution methods. We observe significant differences between STEC and current models, e.g., for 7nm CPU the difference is 13.69%. We further examine the impact of STEC models on existing embodied carbon accounting schemes on computer applications, such as Large Language Model (LLM) training and LLM inference. We observe that using STEC results in much greater differences in the embodied carbon of certain applications as compared to others (e.g., 32.26% vs. 6.35%). This thesis provides comprehensive models to support carbon-intelligent computing, facilitating both operational and embodied carbon optimization across the lifecycle of computers and AI systems. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14417