ABSTRACT
Graph neural networks (GNN) represent an emerging line of DNN algorithms that operate on graph structures. It is becoming more and more popular due to its high accuracy achieved in many graph-related tasks. However, GNN is not as well understood in the system and architecture community as its counterparts such as MLP and CNN. This work tries to introduce the GNN to our community by comprehensively evaluating and analyzing the computation of a set of representative GNN algorithms. The algorithms are selected based on a general GNN computation model and constructed on top of two widely-used GNN libraries. With comprehensive experiments, we make an analysis of GNN computation with respect to general-purpose and application-specific architectures and hope our work can foster more system and architecture research for GNNs.