Imitation Learning-based Fast Optimization of SSD Interface for PCIe 6.0 Considering Signal Integrity

In this paper, we propose an genetic algorithm (GA)-based imitation learning (IL) for fast optimization method of interface in the Peripheral Component Interconnect Express (PCIe) 6.0 system using Pulse Amplitude Modulation (PAM4) signaling. PCIe is the main interface standard for connection between CPU and GPU or solid-state drive (SSD) with a high data rate. However, issues related to signal integrity (SI) become severe as the data rate per lane has increased with each generation. Therefore, optimizing both the PCIe channel and equalizer is required. The proposed method trains the policy network to obtain immediately the optimal design of the channel and equalizer by using the imitation learning method. Based on the high-quality data obtained from the GA, the policy network learns the expert trajectories. Furthermore, the training process is implemented in several host systems. Hence, the trained policy network has reusability for an arbitrary electrical characteristic of the host systems. For verification, the proposed method is applied to the E1.L SSD interface design task. As a result, the superiority of the proposed method is validated by comparing conventional optimization algorithms in terms of optimal performance and computational time.