Hardware Acceleration of Inference on Dynamic GNNs

Dynamic graph neural networks (DGNNs) play a crucial role in applications that require inferencing on graph-structured data, where the connectivity and features of the graph evolve over time. The proposed platform integrates graph neural network (GNN) and recurrent neural network (RNN) components of DGNNs, providing a unified platform that captures spatial and temporal information. Novel contributions include optimized cache reuse, a novel caching policy, and efficient GNN-RNN pipelining. Average energy efficiency gains of 8393X, 183x, and 87X - 10X, and inference speedups of 1796X, 77X, and 21x - 2.4X, over Intel Xeon Gold CPU, NVIDIA V100 GPU, and prior approaches, respectively, are demonstrated across multiple graph datasets and multiple DGNNs.