Neural Architecture Search for In-Memory Computing-Based Deep Learning Accelerators

The rapid growth of artificial intelligence and the increasing complexity of neural network models are driving demand for efficient hardware architectures that can address power-constrained and resource-constrained deployments. In this context, the emergence of in-memory computing (IMC) stands out as a promising technology. For this purpose, several IMC devices, circuits and architectures have been developed. However, the intricate nature of designing, implementing and deploying such architectures necessitates a well-orchestrated toolchain for hardware–software co-design. This toolchain must allow IMC-aware optimizations across the entire stack, encompassing devices, circuits, chips, compilers, software and neural network design. The complexity and sheer size of the design space involved renders manual optimizations impractical. To mitigate these challenges, hardware-aware neural architecture search (HW-NAS) has emerged as a promising approach to accelerate the design of streamlined neural networks tailored for efficient deployment on IMC hardware. This Review illustrates the application of HW-NAS to the specific features of IMC hardware and compares existing optimization frameworks. Ongoing research and unresolved issues are discussed. A roadmap for the evolution of HW-NAS for IMC architectures is proposed. Hardware-aware neural architecture search (HW-NAS) can be used to design efficient in-memory computing (IMC) hardware for deep learning accelerators. This Review discusses methodologies, frameworks, ongoing research, open issues and recommendations, and provides a roadmap for HW-NAS for IMC.