Using Neural Networks for Far-End Crosstalk Compensation in High-Speed MIMO Channels

Abstract

Compared to differential signaling, single-ended signaling is significantly more susceptible to far-end crosstalk; however, single-ended communication is still the preferred data transfer method to use in high-density applications, such as in double data rate (DDR) systems. It has been shown that for loosely coupled single-ended transmission lines, far-end crosstalk (FEXT) on a victim line is proportional to a scaled negative derivative of the far-end signal on the aggressor line; thus, a variety of crosstalk derivative cancelation techniques have been developed. However, when channels are tightly coupled, the derivative cancelation method fails, thus preventing its use at higher data rates. In this article, we examine the derivative-based crosstalk-cancelation technique, and then provide reasons as to why it fails at higher data rates and develop a rule establishing when it can be used. We also propose the use of a time delay neural network crosstalk canceler (NNXC) to cancel FEXT. The proposed crosstalk canceler can operate at significantly higher data rates than cancelers using the derivative-based method. The NNXC can also be used in systems with multiple tightly spaced channels, which is not possible using the derivative method. Furthermore, when a clock signal is available, such as in DDR systems, it can be used as part of the network's training sequence–––significantly improving the performance of the NNXC in reducing far-end crosstalk. Several simulations are shown depicting the superior performance of the NNXC canceler, including in a realistic DDR5 channel with tightly coupled lines.