RX Equalization for a High-Speed Channel based on Bayesian Active Learning using Dropout

Determining optimal equalization settings in highspeed bus design simulations is becoming more important due to increased complexity and data rates of current server systems, but it is also time and resource consuming. In this paper, a probabilistic machine learning technique, Bayesian Active Learning using Dropout (BAL-DO), is utilized to perform RX equalization and optimization to address this issue. Largest HEYE opening and corresponding equalization settings are obtained with high prediction accuracy without performing extensive time-domain analysis, thereby significantly reducing the cost of engineering time and computational resources.