FAME: An Open-source Fast Memory-Efficient Simulation Framework for Non-Ideal ReRAM

Resistive random-access memory (ReRAM) is currently viewed as a potential memory medium for deep neural network (DNN) accelerators within the compute-in-memory paradigm. However, the accuracy of DNN dramatically degrades due to the non-idealities of ReRAM. Efficient non-ideal ReRAM simulator is indispensable for exploring possible optimization space. The existing simulation frameworks face the problems of high time consumption and large GPU memory occupation due to the circuit-level abstraction, which is unacceptable as models become larger. To address the problem, we introduce FAME, a fast data-level simulation framework for non-ideal ReRAM, including comprehensive non-idealities for ReRAM crossbar cells. Proposed S-M-R (Split-Map-Reorganize) strategy transmutes error from non-idealities directly into error of weight data, facilitating data-level simulation in FAME. This strategy effectively diminishes the time and memory overhead. Experimental results show that FAME achieves a speedup of 9.82× and utilizes only 1/4 of the GPU memory compared with prior non-ideal ReRAM-based DNN accelerator simulators. The proposed framework is available in https://github.com/Zhang-XiaoY/FAME.