Optimizing Multi-Dimensional Efficiency of Foundation Models: From Fine-Tuning to Inference

With the rapid advancement of foundation AI models and the widespread adoption of AI applications, the operational energy demand and environmental impacts of AI, have become increasingly significant. Evaluating AI systems should extend beyond accuracy and wall-clock performance to treat sustainability as a first-class objective. However, optimizing sustainability of AI systems is challenging due to the inherently multi-dimensional nature of efficiency (e.g., runtime, memory footprint, and energy) and the fact that operational emissions depend on both energy consumption and time- and location-varying carbon intensity under real-world constraints. This dissertation proposes an end-to-end methodology for sustainable AI that integrates measurement, multi-dimensional evaluation, and decision-making from fine-tuning to deployment scheduling. First, it introduces a multi-dimensional efficiency evaluation framework spanning runtime, GPU memory footprint, and energy consumption, demonstrating why single-metric optimization can be misleading. Second, it develops the first resource-aware decision frameworks that translate multi-dimensional efficiency measurement into actionable configuration choices for sustainable LLM fine-tuning under heterogeneous GPU environments. Third, it presents the first carbon-aware scheduling simulator that evaluates and guides carbon efficient and sustainable decisions under realistic capacity and latency constraints. Together, these contributions provide a systematic foundation for measuring, evaluating, and improving the sustainability of AI workloads in heterogeneous cloud environments.

Name thesis:degree_name

Doctor of Philosophy

Level thesis:degree_level

Doctoral

Discipline thesis:degree_discipline

Computer Science

Grantor

Texas State University

Year dc:date.issued

2026

Author dc:creator

• Chen, Dayuan

Advisor dc:contributor.advisor

• Zong, Ziliang

Committee members dc:contributor.committeemember

• Islam, Tanzima
• Metsis, Vangelis
• Chen, Heping

• cloud computing
• sustainability
• artificial intelligence
• fine-tuning
• optimization
• carbon awareness
• workload shifting

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10877/24848

OAI identifier oai:identifier

oai:digital.library.txst.edu:10877/24848