Machine Learning-based aging compensation to enable safety-critical autonomous driving products

souhir mhira


This paper shows new insights on the stochastic nature of aging-related timing impact in digital circuits. Varying critical paths through aging trigger the need for aging compensation control loop based on in-situ slack monitors (IS2M) to cope with intradie variations. A new aging compensation mechanism is proposed based on deterministic patterns run periodically during Power-On sequences. The number of pre-error flags being stochastic, it is mandatory to introduce an unsupervised machine learning algorithm to control supply voltage changes. Adaptive Resonance Theory (ART) algorithm is favored for its ability to handle the stability-plasticity dilemma. This work clearly states the design needs, theoretical background and the experimental validation of such approach so to enable safety-critical autonomous driving products in advanced CMOS nodes.