Fast Channel Analysis and Design Approach using Deep Learning Algorithm for 112Gbs HSI Signal Routing Optimization

This work proposes a novel deep learning-based design approach that performs fast and flexible SI analysis on high-speed interface (HSI) channel structures with various impedance mismatching sections. The concept of the proposed method is to model segments (i.e., unit-length differential traces and differential vias) which compose a channel structure to be analyzed, and then to complete the entire channel modeling by combining the generated deep learning (DL)-based segment models together. In this approach, deep neural networks (DNNs) are used to model frequency responses (S-parameter) of unit-length differential traces and differential vias by using their geometric parameters as input. After achieving DL-based models of unit-length differential traces and differential vias from the DNNs, the proposed method applies cascading operations to the generated S-parameter models of the segments to create the entire frequency response of the desired channel structure. We conducted an experimental study to evaluate the effectiveness of the proposed DL-based channel modeling approach. The experimental results demonstrates that the proposed approach has very high modeling accuracy with R-squared value > 0.9 on average. Furthermore, it is found that the modeling speed of the proposed method is more than 1,000 times faster than that of conventional physics-based EM solvers. Therefore, the proposed method can be used to boost efficiency of SI analysis and, also, support effective design space exploration and optimization of HSI channel structures for the latest SerDes solutions.