An Efficient Optical Sparse Matrix Multiplication Accelerator for Graph Neural Networks

Graph Neural Network (GNN) is a variant of Deep Neural Networks on graphs and have been widely used in the real world. However, GNNs are more complex compared to traditional DNNs due to the unstructured and sparse graph data. Existing optical computing platforms do not support sparse operations, requiring a large number of invalid zero mapping to the computing unit to maintain the correct value of the final results. Therefore, in this paper, we propose an optical graph neural network (OGNN), a novel sparse matrix multiplication accelerator that can implement GNN to address this problem. OGNN features two key techniques: reconfigurable sparse matrix-vector multiplication (SpMV) operation and graph partitioning operation, which uses temporal, wavelength, and spatial dimensions to construct an optical parallel processing framework for graph neural network. We evaluate OGNN with two graphs datasets - Cora and PubMed. The results show that OGNN achieves competitive performance with the electronic counterpart and provides efficient computing unit utilization for sparse multiplication.