Graph neural network based on RRAM array

Graph Neural Network (GNN) has emerged as a powerful method for processing graph data structure promising in a large variety of applications, such as social networks, natural science and knowledge graphs. Meanwhile, in-memory computing has great potential in neural network, especially in energy consumption and speed issues. In this study, we first quantize GraphSAGE and then implement both inference utilizing device 4-level characteristic and online training utilizing LTP<D characteristic. The implementation of quantized GraphSAGE algorithm accelerated by Pt/Ta/Ta2O5/Pt/Ti RRAM array is simulated. Through simulation experiments, we demonstrate the advantages and prospects of GNN based on RRAM array.