ABSTRACT Graph processing is used in many domains to extract knowledge from real‐world data. With the rise of deep neural networks and scaled compute infrastructure in artificial intelligence (AI), specialized techniques emerged to leverage graphs in applications such as recommendation systems and social networks. Numerous graph neural network (GNN) models and frameworks have been developed, extending the capabilities of graph analytics. While much effort has focused on improving model accuracy, less attention has been given to analyzing resource usage during GNN processing. This work presents a quantitative analysis of resource consumption during GNN model training and examines the benefits of GPU acceleration across datasets with different structural properties. In the largest workload evaluated, GraphSAGE trained on the coauthors‐CS dataset achieved a 27.6× speedup in the GPU‐accelerated environment compared to CPU‐only execution, representing an order‐of‐magnitude improvement.
Ribeiro et al. (Fri,) studied this question.