A study on Machine Learning-based Hardware Bug Localization

Abstract

Simulation-based verification is a very essential technique in ensuring the correct functionality of any digital integrated circuit design before it goes on silicon. One of the major challenges of running simulation-based verification on complex designs is the tradeoff between simulation time and the time taken for failure localization or to root cause. This is because the simulation run times could be very high when there are many checkers used per cycle of execution. However, when lesser checkers are turned on, the amount of time for manual debug increases because, after failure, the verification engineer has to manually analyze the failure and turn on the more granular checkers individually and re-simulate; or invest lots of time, memory and resources to manually go through the simulation cycles dumps before the failure which is not good given the current complexity of designs.

Machine learning has emerged to be a popular technique to construct mathematical models that can understand the expected patterns from a given dataset. To address the aforementioned trade-off problem, an idea is investigated to use the failing signatures from fewer active high-level checkers during simulation to train a machine learning model to predict the location of the bug in the design. This information would in turn be used to turn on relevant checkers in the design before re-simulation. Other methods to analyze the signals in design after failure to predict bug location were also studied. This idea is implemented and tested on a MIPS processor with total of ~ 700 bugs injected in 15 different units to distinguish them with good accuracy.