MemNAS: Super-net Neural Architecture Search for Memristor-based DNN Accelerators

Processing-in-memory (PiM) is an emerging technology that is particularly attractive for deep neural networks (DNN), providing a reduced power consumption at the price of non-idealities in the computations. Furthermore, the PiM platform offers additional parameters to optimize power consumption and non-idealities' effects on computations. However, PiM parameters may need to be tuned differently for each DNN model. In this work, we consider jointly optimizing the DNN architecture and some PiM parameters to improve the Pareto front of DNN performance and power consumption. This paper reports the results of neural architecture search (NAS) experiments on a Once-for-all (OFA) ResNet50 on ImageNet1K and finds that more power efficient DNN models can be identified with the same computational cost as the original OFA method.