Inference of Deep Neural Networks with Analog Memory Devices

Abstract

Acceleration of Deep Neural Networks (DNNs) inference using non-volatile memory (NVM) arrays, such as PhaseChange Memory (PCM), shows promising advantages with respect to digital implementations, in terms of energy efficiency and speed. By exploiting closed-loop tuning techniques and weight encoding using multiple conductances with varying significance, we experimentally demonstrate, using up to 220 PCM devices, careful analog conductance tuning. Such programming techniques are implementable in a row-by-row scheme, enabling reliable weight encoding for future inference hardware.