Architectural Implication of Graph Neural Networks

Abstract

Graph neural networks (GNN) represent an emerging line of deep learning\nmodels that operate on graph structures. It is becoming more and more popular\ndue to its high accuracy achieved in many graph-related tasks. However, GNN is\nnot as well understood in the system and architecture community as its\ncounterparts such as multi-layer perceptrons and convolutional neural networks.\nThis work tries to introduce the GNN to our community. In contrast to prior\nwork that only presents characterizations of GCNs, our work covers a large\nportion of the varieties for GNN workloads based on a general GNN description\nframework. By constructing the models on top of two widely-used libraries, we\ncharacterize the GNN computation at inference stage concerning general-purpose\nand application-specific architectures and hope our work can foster more system\nand architecture research for GNNs.\n