PCIe Gen6 Physical Layer Equalization Tuning by Using Unsupervised and Supervised Machine Learning Techniques

Ever faster applications triggered the development of the new PCIe Gen6 specification, reaching 64 GT/s data rates with PAM4 modulation. This brings new challenges for the physical channel design, where equalization (EQ) plays a key role. PCIe specification defines an EQ process at the transmitter (Tx) and the receiver (Rx). Current post-silicon validation practices consist of finding optimal subsets of Tx and Rx coefficients by measuring the eye diagram at the Rx across many different channels. However, these practices are very time consuming since they require massive lab measurements. In this paper, we propose machine learning techniques to cluster post-silicon data from different channels and feed those clusters to Gaussian process regression (GPR) models. We then optimize each GPR surrogate to obtain the optimal tuning settings for each identified cluster. Our methodology is validated by using MATLAB SerDes Toolbox simulations of the functional eye diagram of a Gen6 link.