SynAct: Adaptive LLM Agent for Logic Synthesis Optimization

A team of researchers has developed SynAct, an adaptive closed-loop reasoning-acting agent based on a large language model (LLM), aimed at enhancing logic synthesis—a vital phase in chip design that converts register-transfer level (RTL) designs into gate-level netlists. The effectiveness of the resulting power, performance, and area (PPA) metrics is highly influenced by the selection of optimization commands, making the tuning process complex and costly. Unlike traditional automated techniques that rely on fixed action spaces with limited interpretability, and existing LLM methods that create static scripts, SynAct overcomes these challenges by continuously analyzing live synthesis reports and reasoning based on the current circuit state, historical optimization data, and retrieved tool knowledge to issue specific commands. Its focus is on improving timing, especially worst negative slack (WNS), while balancing area and power. Tests on a commercial synthesis tool across 14 designs indicate that SynAct significantly enhances timing metrics compared to standard methods. This system's capability to adapt in real-time to changing circuit conditions marks a significant step forward in utilizing LLMs for electronic design automation, providing a more intelligent and responsive synthesis optimization strategy.