Fig. 2: Overview of KAN’s hardware architecture strategies.

a Reference object and benchmark metrics for the hardware properties of KAN. b The bottlenecks encountered by the von Neumann and near-memory architectures in processing the computations of KAN, while the GMC-based architecture provides a reliable solution at the in-memory computing level. c The GMC ingeniously combines the Gaussian kernel provided by AAT with the parameterization capability of memristors, integrating the necessary physical foundation for the in-memory computing required by KAN.