Graph Neural Networks (GNNs) use a fully-connected layer to extract features\nfrom the nodes of a graph and aggregate these features using message passing\nbetween nodes, combining two distinct computational patterns: dense, regular\ncomputations and sparse, irregular computations.\n To address this challenge, we propose GNNerator, an accelerator with\nheterogeneous compute engines optimized for these two patterns.\n Further, GNNerator implements feature-blocking, a novel GNN dataflow that\nbeneficially trades off irregular memory accesses during aggregation for\nregular memory accesses during feature extraction. We show GNNerator achieves\nspeedups of 5.7-37x over an NVIDIA RTX 2080-Ti, and 2.3x-3.8x over HyGCN, a\nstate-of-the-art GNN accelerator.\n