A Diffusion-Based Framework for Decoupling Capacitor Placement Optimization

The problem of decoupling capacitor placement is a critical challenge in modern hardware design, as it significantly impacts power integrity and overall circuit performance. The rapid evolution of hardware to meet increasing performance demands presents significant challenges for traditional Electronic Design Automation (EDA) tools, which often suffer from prolonged runtimes and limited scalability. Recently, sequential decision-making methods have gained attention for automating partial solutions in hardware design. However, their dependence on a strict ordering of decision elements introduces order bias, which may degrade performance and overlook valid solutions when the optimal solution set is permutation-invariant. In this paper, we propose a diffusion-based framework for the problem of decoupling capacitor placement, leveraging its permutation-invariant nature to address these challenges. Unlike sequential decision-making methods that introduce order bias, our approach is the first to leverage a non-autoregressive diffusion model for this task, directly generating an optimized placement solution by iteratively refining a noisy initialization. This allows the model to efficiently explore the solution space while ensuring consistency and high-quality placements. Our experiments demonstrate that our method outperforms multiple state-of-the-art (SOTA) baselines, including search-based heuristics, reinforcement learning, and imitation learning, in both placement performance and data efficiency.