Endurance-Aware Deep Neural Network Real-Time Scheduling on ReRAM Accelerators
TimeTuesday, July 12th6pm - 7pm PDT
LocationLevel 2 Lobby
Event Type
Networking Reception
Work-in-Progress Poster
DescriptionResistive Random-Access Memory (ReRAM) exhibits a high data density and a low computational cost by in-situ processing, but its endurance is magnitudes lower than traditional DRAMs. This work proposes an Endurance-Aware Real-Time DNN Scheduling (EAS) strategy to enhance the ReRAM endurance under latency constraints. First, a pre-processing methodology is proposed to transform a DNN to an end-to-end execution sequence for resource partitioning. Then, a periodic real-time scheduling method is developed to extend ReRAM programming cycles without violating deadline constraints. The experiment results show that our EAS approach can extend the baseline ReRAM lifetime more than 3 times on average, at a computational cost of less than 1ms.