Reinforcement Learning-Based via Placement Optimization in Package Substrate for Multiple Power Domain 3D-ICs

In this paper, reinforcement learning-based via placement optimization methodology is proposed for multiple power domain (MPD) 3D-ICs. The package substrate is responsible for routing power and ground to appropriate power domains and components in MPD 3D-ICs. Proper design of the package substrate contributes to obtaining power integrity by efficiently allocating power resources and lowering PDN impedance. Among components in the package substrate, microvia transfers power vertically and lowers PDN impedance through proper assignment. However, the package substrate is exposed to the optimization challenges resulted from MPD. Accordingly, a reinforcement learning algorithm is adopted with a deep learning network to optimize the power/ground microvia placement so that power noise can be minimized. Markov decision process (MDP) is defined and simultaneous switching noise (SSN) is modeled with analytical modeling of the package substrate component. The proposed method demonstrates the superior optimality performance compared to meta-heuristic optimization algorithms while consuming lower execution time.