Endurance-Aware Deep Neural Network Real-Time Scheduling on ReRAM Accelerators

Achieving accurate multi-modal Deep Neural Networks (DNN) testing often requires operating rich model parameters under limited computing and memory resources. The low-cost Resistive Random-Access Memory (ReRAM)-based DNN accelerator is a promising solution for such an application thanks to its inherent processing-in-memory capability. However, its lifetime, which is crucial for applications with strict reliability standards such as self-driving cars, can be significantly limited due to: 1) the need to frequently switch weights among different models for real-time streaming applications; 2) the low endurance of ReRAM devices compare to DRAMs by orders of magnitude. This work proposed an Endurance-Aware multi-modal DNN Scheduling (EAS) strategy to address this issue using real-time techniques. First, a pre-processing methodology transformed a DNN into an end-to-end execution sequence for partitioning and scheduling. Then, a periodic real-time scheduling method was developed via data reusing for extending ReRAM programming cycles under deadline constraints. The experiment results showed that our EAS approach can extend the baseline ReRAM accelerator's lifetime from 0.98 years to 3.14 years on average at a low scheduling overhead.