Secure and reliable in-memory computing for edge AI

Abstract

Security measures that are portable, energy-efficient, and attack-resistant are desperately needed given the explosive growth of Edge Artificial Intelligence (Edge AI) systems. Conventional encryption techniques are not appropriate for edge devices with limited resources since they frequently need a significant computational cost. Using the inherent characteristics of 8T Static Random Access Memory (SRAM) cells and the bitline discharge technique, this thesis explores a secure and reliable In-memory computing architecture to implement Physical Unclonable Functions (PUF) and True Random Number Generators (TRNG) directly within memory arrays. The major goal of this work is to create and evaluate a comprehensive 8T SRAM-based framework that can handle secure key generation based on PUF with little overhead. A single 8T SRAM cell is designed and characterized at the start of the study, which then moves on to a 64×32 array layout. In order to ensure proper operation and performance optimization of the memory array and for PUF implementation, extensive efforts were made to create, integrate, and simulate crucial peripheral circuits such as write driver, precharge circuit, and sense amplifier.