Thermal Induced Retention Degradation of RRAM-based Neuromorphic Computing Chips

The retention characteristic of RRAM devices will obviously decrease the computing reliability of RRAM-based neuromorphic computing chips. In this paper, with the help of thermal simulation, we evaluated the thermal effect on the retention characteristic of RRAM devices and the computing accuracy of RRAM-based neuromorphic computing chips. We find weight dual allocation can achieve $\boldsymbol{ < 1\%}$ accuracy loss in ten years, as compared to 1% accuracy loss in just 4.6 days when one 1T1R unit is used to express one weight. The results also show that the lower temperature can also help to improve the computing reliability of neuromorphic computing chips.