New Technique Combines Resistive Memory with Graph Neural Networks

Discovery could improve efficiency in neural networks by eliminating the von Neumann bottleneck.

A team of scientists at University of Hong Kong, InnoHK, the Chinese Academy of Sciences, and other institutions have developed a method for using resistive memory with a Graph Neural Network (GNN). This technique, which demonstrates that it is possible to maintain the efficiency of resistive memory without incurring the limitation of manual resistive memory programming, could one day lead to more efficient drug design and recommendation systems.

As memory and computer processor chips become more efficient, the system bus separating memory from the processor becomes a limiting factor. This limitation is known as the von-Neumann bottleneck. Resistive memory cells can store data and also perform calculations, which has the potential to eliminate the von-Neumann bottleneck, however, conventional resistive memory is not accurate enough for the graph-level of accuracy used with GNNs. The research focused on developing more accurate resistive memory and tailoring it to meet the needs of the GNN context.

For more information on the study, see the report at HPC Wire, or check out the original research paper published in Nature Machine Intelligence.