Impact of Multilevel Retention Characteristics on RRAM based DNN Inference Engine

In this work, the retention characteristics of multilevel HfO
₂
resistive random access memory (RRAM) based synaptic array was statistically measured from a 90 nm test chip and modeled at different temperatures. We found that not only the average conductance (especially at the intermediate states) drifts but also the variance of conductance exacerbates at elevated temperatures. To investigate the impact of the synaptic weight drift on deep neural network, the experimental data are modeled into the ResNet-18 simulation with 1-4 weight bit precisions. The result shows that the inference accuracy drops significantly at 55°C or above, which implies further engineering on RRAM retention or circuit/algorithmic techniques are yet to be applied.