Abstract

Deep learning workloads generate substantial carbon emissions in data centers, driven by long training durations and intensive energy use. To address this challenge, previous studies have explored temporal workload shifting and spatial workload migration. Yet, these approaches remain limited for long-running workloads, such as large language models, because they fail to adapt to continuous fluctuations in regional carbon intensity. We introduce GreenAccounter, a toolkit for carbon-aware orchestration of deep learning workloads across multi-cloud environments. It integrates real-time carbon intensity monitoring with checkpoint-based migration, allowing training to continue seamlessly while reducing emissions. A unified dashboard visualizes regional carbon intensity, cumulative emissions, and power consumption, providing operators with a single pane of glass for managing distributed cloud resources. GreenAccounter functions both as a reproducible research platform for carbon-aware scheduling and as a practical operational toolkit for emission reduction in AI training. Released as open source, it promotes sustainable, transparent, and data-driven practices for deep learning at scale.