Applying Multi-objective Acquisition Function Ensemble for a candidate proposal algorithm

Pre-silicon verification takes up to 70% of the circuit development time. Thus, it constitutes a significant bottleneck in the production of new circuit designs. Taking this into consideration, our current paper proposes new developments to an existing machine learning based verification algorithm, by integrating a multiobjective acquisition function ensemble (MACE) step, to be used as a candidate selection mechanism. We report results obtained on both synthetic and real circuits. Also, we compare the improved algorithm with the previous version, which used gradient descent (GD) to choose candidates for our next simulation, in order to identify possible circuit failures. By candidates, we refer to the set of input operating conditions (OCs) of the circuit. One functional difference, between MACE and GD, is that MACE uses 3 objective functions to calculate a score for each candidate, whereas GD uses just one. We demonstrate that the improved algorithm performs better on the synthetic circuit. Moreover, it offers better candidates for the majority of the real circuit responses. Furthermore, the MACE variant was able to spot an operating condition combination (OCC), that led to specification violation. It is worth noting that the GD approach failed to find this particular OCC. Therefore, MACE showed great potential as an improvement to the already established algorithm.