Quantifying CO2 Emission Reduction Through Spatial Partitioning in Deep Learning Recommendation System Workloads
Andrei Bersatti, Euna Kim, Hyesoon Kim · IEEE Micro · 2024
The resource demand of modern applications has been increasing at a dizzying pace. This rising prominence is accompanied by a rising concern for the greenhouse gas emissions and carbon footprints of these applications. Deep learning recommendation models (DLRMs) are one example of the important models that are rising in importance and ubiquity in data centers. In this work, we analyze the impact of spatial partitioning of workloads that can decouple the geographical constraint and thus achieve reduced CO2 emissions by using DLRM online training as the workload under study.