Optimization of the Power/Ground Ball Map in 3-D-IC BGA Packages With Multiple Power Domains Using Deep Reinforcement Learning

Abstract

In this article, a deep reinforcement learning (DRL) approach that optimizes the power/ground ball map design within the ball grid array (BGA) package of 3-D-integrated circuits (ICs) considering multiple-power-domain (MPD) environments is proposed. With the adoption of the MPD in 3-D-ICs, the BGA in the package substrate, which routes the power and ground to the appropriate power domain and components from the board, faces design challenges such as highly complex power noise between the domains and allocations considering multiple power densities. To address this design complexity, a U-net-based DRL method is adopted, which trains an optimal policy to obtain an optimal power/ground ball map to mitigate simultaneous switching noise (SSN). The U-net enables the conservation of semantic and spatial information simultaneously, allowing the proposed method to optimize the complex ball map effectively within the BGA package for MPD. Furthermore, the proposed method shows reusability and scalability by training with variously displaced current source center (CSC) balls and various ball map sizes while maintaining the power grouping design. The optimality performance of the proposed method is verified by comparing it to conventional optimization algorithms such as random search (RS) and a genetic algorithm (GA). The proposed method outperforms RS and GA in terms of execution time and optimality performance across various tests with different ball map sizes and CSC positions.