Getting more out of GPUs for Graph Neural Networks

Graph Neural Networks (GNNs) are an important tool for extracting value from relational and unstructured data. GNNs pose several computational challenges that are not present in neural networks for structured data such as images or sequences. While GPUs provide an ideal environment for processing neural networks due to their high-memory bandwidth and massive parallelism, bottlenecks on network bandwidth, the CPU, and PCIe bandwidth have lead to low GPU utilization during GNN training and inference. In this talk, we look at various techniques for minimizing data movement across the network and PCIe, off-loading work from the CPU, and improving throughput on GPUs. We demonstrate the effects of these techniques in the popular Deep Graph Library and PyTorch Geometric GNN frameworks