EcoPhaze: A Hierarchical Approach to Carbon-Aware Accelerator Pruning for Distributed Deep Learning Systems
Author(s)
Potluri, Aditya Hari
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
This paper introduces EcoPhaze, a hierarchical approach to carbon-aware accelerator pruning for distributed deep learning systems. Recognizing the escalating environmental impact of large-scale AI, the research addresses the co-optimization of accelerator architecture and distributed training deployment with the explicit goal of reducing carbon footprint and energy consumption. Building upon the Phaze framework for hardware architecture search, EcoPhaze integrates carbon modeling that accounts for both embodied and operational emissions. A novel hierarchical pruning algorithm is proposed, which optimizes for a performance-per-carbon metric, guiding the search towards more sustainable and efficient accelerator designs. Evaluation on BERT-Large demonstrates that EcoPhaze can achieve comparable performance to baseline methods with significantly reduced resource utilization and embodied carbon, highlighting the potential of co-designing hardware and deployment strategies with environmental considerations at the forefront.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis