Abstract
The energy and latency budgets of edge artificial intelligence (AI) are far tighter than in data-centre computing, yet the workloads are no less demanding. Conventional von Neumann hardware spends most of its energy shuttling data between separated memory and logic, a cost the edge cannot absorb. Memristive devices collapse this separation: the same nanoscale element that stores a synaptic weight also performs computation in place, governed directly by the physics of resistive switching. This review follows a single intellectual thread, from the device physics of resistive switching, through the crossbar matrix–vector-multiplication primitive, to in-memory and neuromorphic architectures and, finally, to the algorithmic and deployment constraints unique to the edge. We argue that the edge is the regime where device physics most directly dictates system capability—a case study in the pervasiveness of physics across engineering and computing. We compare filamentary, interface-type and threshold-switching mechanisms by the computing role each enables; survey analog in-memory, analog/digital-fusion, near-memory and event-driven neuromorphic architectures; and assess recent chip demonstrations through reported energy efficiency (TOPS/W) and accuracy-after-hardware-mapping, read against the digital edge baseline. We close with applications in biomedical sensing and hardware security, and the co-design challenges separating prototypes from deployable systems.