Information for Paper ID 7008
Paper Information:
Paper Title: Dynamically Optimized Test Generation Using Machine Learning 
Affiliation Type: Industry 
Keywords: hardware verification, coverage closure, test generation, constraints optimization, machine learning, bayesian optimization, gradient boosted trees 
Abstract: Modern constrained random testbenches for hardware verification often require engineers to manually optimize test randomization constraints to target different use cases such as feature bringup, bug hunting and coverage closure. This paper presents a practical solution that automates the process of optimizing constrained random test generation by monitoring feedback from the design under test (DUT) and dynamically optimizing randomization constraints. We leverage Bayesian optimization (BayesOpt) [1] with the gradient boosted regression trees (GBRT) machine learning model [2] to optimize constraints in (1) batched offline flows and (2) live online flows. Experimental results in various testbenches demonstrate that the offline flow can boost coverage counts across coverpoints in a design by up to 6x and improve an interface packet frequency 6.3x. The online flow is shown to improve the occupancy of a complex FIFO in the design by 7x. 
Track ID: 3.1 
Track Name: Automating the Optimization of Verification Processes 
Final Decision: Accept as Lecture 
Session Name: Automation using Machine Learning (Lecture) 
Author Questions:
Confirmed: Yes