Reinforcement Learning-Based Decap Optimization Method for High-Performance Solid-State Drive

In this paper, we propose an improved optimal decoupling capacitor (decap) design method based on Q-learning algorithm for high-performance solid-state drive (SSD). The proposed method selects optimal decap combinations that satisfies target impedance with minimum decap number. Based on Q-learning algorithm combined with transmission line theory, optimal decap combinations of power distribution network (PDN) can be provided. The proposed method was verified with voltage ripple measurement and PDN impedance simulation using SSD for high-performance server application. Conventional decap optimization method are using complex and time-consuming analytical tool with power integrity (PI) domain expertise. However, the proposed method requires only the PDN and decap information along with a simple Q-learning model without PI knowledge, providing faster and accurate results than full search optimization method. For example, in 21 decaps combination problem, the proposed method’s computing time consumes only few minutes, 89.09 sec, which is significantly reduced result compared with the conventional full search simulation. Therefore, we expected the proposed method can be widely used to solve for decap optimization problem with complex PDN.