Large Language Model-Based Decoupling Capacitor Placement Optimization

In this paper, for the first time, we propose a large language model (LLM)-based decoupling capacitor (decap) placement optimization method. By leveraging Optimization by PROmpting (OPRO), we harness pre-trained LLMs' patternrecognition capabilities, demonstrating their potential to address complex power integrity (PI) domain engineering problems beyond simple question-and-answer interactions. We structured the LLM's input prompts to include task descriptions, domain knowledge, output formats, and information from previous attempts, enabling effective optimization. We applied the proposed algorithm to solve the decap placement problem for power distribution networks (PDNs) composed of high bandwidth memory (HBM) base die, silicon interposer, and package. The proposed approach is compared with traditional optimization algorithms, such as random search and simulated annealing, showing that an LLMbased optimizer can effectively capture PI-related patterns and deliver competitive solutions. This work highlights the potential of LLM-driven optimization in solving complex engineering problems, extending the application of LLMs beyond their conventional use as chatbots or general-purpose assistants.