Memristive In-Memory Computing for AI: Neural Network Acceleration, Adaptive Learning, and System Integration

Abstract

Moving data between memory and processing units has become a major cost in neural network computing. Memristive crossbars address this cost by retaining model parameters as conductance states and performing parallel multiplication and accumulation within the array. This review examines how that physical operation is translated into numerical formats, peripheral circuits, neural-network mappings, heterogeneous accelerators, local learning, compilation, and calibration. Fabricated chips have demonstrated substantial parallelism and lower weight traffic. Their application performance, however, is determined just as strongly by utilization, data conversion, unsupported operations, programming overhead, and device variation. Representative heterogeneous systems combine memristive arrays with static random-access memory (SRAM), digital logic, and software, assigning each resource the work it handles most efficiently. Current hardware evidence is strongest for inference and limited adaptation. Continuous monitoring, remapping, and recovery remain largely runtime design goals. Peak array efficiency alone is therefore insufficient; comparisons should use measured workloads and clearly stated system boundaries.