Optimizing Constrained Random Verification with ML and Bayesian Estimation

Constrained Random Verification (CRV) is a widely used design verification methodology that focuses on generating pseudo-random legal transactions or stimuli for the device under test (DUT) to weed out corner cases or hidden bugs that cannot be easily anticipated. However, it is limited by the difficulty of determining when to cease generating stimulus patterns. In this paper, we describe a novel approach that integrates machine learning and Bayesian estimation to estimate the number of times a specific random test-bench must be executed to identify all failures within a budget with high accuracy. We evaluate this method on a collection of CRV flows on real-world industrial designs during development and demonstrate up-to 86.9% reduction in test instances with 99.2% accuracy to detect the same failures as baseline estimate-oblivious methods.