Multi-Fidelity Nonideality Simulation and Evaluation Framework for Resistive Neuromorphic Computing

While ReRAM crossbar arrays (RCAs) can perform matrix-vector multiplication (MVM) operations with great efficiency, they often suffer functional inaccuracy due to hardware nonidealities, making the functional evaluation of RCA-based systems a central challenge for system designers. Along with a taxonomy of RCA modeling approaches, this paper presents NIXSim, a nonideality-aware simulation and evaluation frame-work for RCA-based neural network accelerators with an em-phasis on accuracy and speed tradeoff. We also present crossbar-and network-level comparisons with previous work, and offline training examples demonstrating the usefulness of our framework for designing RCA-based neural network accelerators.
^1
1
Code is available at https://github.com/QCH170308/NIXSim.