Task 004 - Neuromorphic Fabrics

Bio: Jacob R. Stevens is a PhD candidate in the School of Electrical and Computer Engineering at Purdue University. He received his B.S in Computer Engineering from Purdue as well. His current research interests include architectural and algorithmic approaches to accelerating neural networks, particularly emerging neural workloads such as memory-augmented and graph neural networks. 

 

Abstract: Graph Neural Networks (GNNs) are a promising new class of Deep Neural Networks (NNs) that aim to bring the success of deep learning from the Euclidean domain to the non-Euclidean domain of graphs. GNNs extract features from the nodes (or edges) of a graph using dense, regular computations, and aggregate these features using message passing between the nodes (edges), resulting in sparse, irregular computation. Current computing platforms do not adequately address this combination of compute pattern. Thus, we propose GNNerator, a graph neural network accelerator that utilizes two heterogeneous compute engines optimized for the differing needs of feature extraction and feature aggregation. Further, GNNerator implements feature-blocking, a novel dataflow for GNNs that helps to alleviate the memory bottleneck inherent in GNNs. Across a suite of different graph datasets and networks, GNNerator achieves speedups of 5.7-37x over an NVIDIA RTX 2080-Ti GPU, and 1.8x-8x over HyGCN, a recently proposed GNN accelerator.