Next-generation graph computing with electric current-based approaches (www.nature.com)

A new Perspective outlines next-generation hardware approaches for graph computing, centering on electric current–based graph computing (EGC) built from memristive crossbar arrays (CBAs) and a class of quantum‑inspired graph computing (QGC) methods such as probabilistic bits (p‑bits), oscillatory neural networks (ONNs), and Hopfield neural networks (HNNs). The paper highlights recent advances that move EGC beyond simple Euclidean graphs: CBAs can now encode directed and weighted non‑Euclidean graphs, multi‑CBA setups model probabilistic connectivity, and self‑rectifying memristors address directional scaling constraints. QGC leverages two‑level and phase‑based nodes to implement energy‑based optimization (Ising‑style) without requiring coherent qubits, using p‑bits that stochastically toggle between 0 and 1, ONNs that exploit phase coupling, and HNNs that evolve via energy gradients. This work is significant because it reframes graphs as physical substrates—currents, resistive states, or oscillatory phases—enabling massively parallel, low‑latency solutions for connectivity, probabilistic inference, and combinatorial optimization central to ML, biology, materials, and social-data problems. Key technical challenges remain: CBAs face sneak currents and require high endurance, symmetric weight updates, and directional device properties; QGC needs robust device primitives and architectures to translate Ising mappings into reliable hardware. The Perspective argues that advancing materials, selectors, 3D integration, and hybrid architectures will be crucial to realize scalable, energy‑efficient graph processors that complement classical and quantum approaches.