ATCPiM: Analog to Time Coded Processing in Memory for IoT at the Edge

Abstract

Data conversion from analog to digital and time domains and data transfer costs are the main bottlenecks in designing energy & area efficient near-sensor processing architectures. In this paper, we propose a temporal Processing in Memory architecture that addresses both data conversion and data transfer costs. We develop a novel analog to time encoding method that translates the incoming analog signal directly into a time-coded value, reducing data conversion costs. To do this, we exploit the linear kinematics of skyrmions in magnetic racetracks and propose a skyrmion-based non-volatile temporal memory architecture that employs time encoding. We also reduce the data transfer costs by processing the temporally stored data during the memory read operation. A read unit shared between the temporal memory cells acts as the processing element (PE) in our proposed design. Thus, the proposed approach enables the reduction of both data conversion and data transfer costs achieving < 340 fJ/conversion and approximately 530-660 fJ/operation.