Degradation analysis and optimization of temperature effect on MEMRISTOR-based Neural Network Accelerators by electro-thermal simulation

Nowadays, memristor-based neural network accelerators have been widely studied due to their outstanding performance in massive parallel vector matrix multiplication. However, the memristor is sensitive to temperature and its on/off state operation window can be seriously degraded by the increasing temperature, which may lead to computation failures in memristor -based NN accelerators. In this work, we establish an electro-thermal simulation platform to evaluate the temperature impact on memristor -based NN accelerators. With this platform, we first investigate the impact on computation accuracy with the temperature increase in different NN layers in the accelerators. We then apply a temperature-aware NN weight mapping scheme to the most temperature-sensitive layer and achieve 28.89% improvement in computation accuracy, which only has 0.06% difference with the improvement achieved by applied the mapping scheme to the whole NN model. This finding can help to simplify the temperature-aware hardware optimization design in memristor-based neural network accelerators and reduce the power consumption.