abstract
- The rapid increase of computationally intensive workloads, particularly in machine learning and large-scale data processing, has led to a substantial rise in data center energy consumption, contributing to concerns about the global sustainability of computing. Graphics Processing Units (GPUs) are central to the infrastructure, with each new generation offering higher performance and increasing energy efficiency. However, these gains come with an embodied carbon footprint from manufacturing, which can outweigh operational savings depending on workload, utilization, and grid carbon intensity. In this work, we conduct a lifecycle analysis of server-grade GPUs released over the past decade, estimating embodied carbon from manufacturing and operational emissions across varying workloads, deployment locations, and scaling factors. Using break-even analysis, we identify whether GPU upgrades yield net ecological benefits, as well as scenarios where they do not. Our results provide guidance for data center operators and GPU users, supporting sustainability-aware hardware procurement and informed timing of GPU replacements.