Deep learning software stacks for analogue in-memory computing-based accelerators

Analogue in-memory computing (AIMC) is an emerging computational paradigm that can efficiently accelerate the key operations in deep learning (DL) inference workloads. Heterogeneous architectures, which integrate both AIMC tiles and digital processing units, have been proposed to enable the end-to-end execution of various deep neural network models. However, developing a software stack for these architectures is challenging, owing to their distinct characteristics — such as the need for extensive or complete weight stationarity and pipelined execution across layers, if maximum performance is to be achieved. Moreover, AIMC tiles are inherently stochastic and hence introduce a combination of stochastic and deterministic noise, which adversely affects accuracy. As a result, existing tools for software stack development are not directly applicable. In this Perspective, we give an overview of the key attributes of DL software stacks and AIMC-based accelerators, outline the challenges associated with designing DL software stacks for AIMC-based accelerators and present opportunities for future research. Analogue in-memory computing (AIMC), with digital processing, forms a useful architecture for performant end-to-end execution of deep neural network models, but requires the development of sophisticated software stacks. This Perspective outlines the challenges in designing deep learning software stacks for AIMC-based accelerators, and suggests directions for future research.