Industry Track: MAMBO: ML Accelerated Multi-Agent Black-Box Optimization for Enhancing Standard Cell Synthesis Quality

Modern VLSI chip design flows necessitate efficient exploration of design space and precise analysis of Performance, Power and Area (PPA), as technology nodes continue to scale down. Electronic design automation (EDA) tools streamline the physical design process, enabling efficient generation, optimization and verification of circuit designs, thereby eliminating the need for manual interventions and facilitating design technology co-optimization (DTCO). However, physical design layout generation often involves an extremely large search space. To be practical, EDA tools need to be guided by hyperparameters. Choosing the right hyperparameter values is challenging since the impact on design space exploration and thus on layout synthesis results is not trivial. This motivates a need for efficient algorithms to explore design space and optimize hyperparameters. In this work, we have developed MAMBO, a synergistic, multi-agent framework that optimizes design parameters and enhances circuit design outcomes. MAMBO integrates multiple ML algorithms, such as Bayesian optimization and reinforcement learning, to work collaboratively with a shared optimization history, enabling rapid discovery of optimal solutions while effectively exploring the search space. We apply MAMBO to optimize the hyperparameters of SLiC™, an automated standard cell layout generation tool from Applied Materials® that efficiently produces DRC- and LVS-correct optimized layouts directly from transistor netlists. We demonstrate that MAMBO's collective intelligence yields enhanced synthesis outcomes compared to SLiC™ results obtained with default parameters or those achieved by a single optimization agent.