Machine Learning-based BER Estimation in Receiver IBIS-AMI Modeling
TimeMonday, July 11th3:45pm - 4pm PDT
Location2012, Level 2
Event Type
Engineering Tracks
DescriptionIt is difficult to make the IBIS-AMI model highly correlated with real circuit behavior. As the operating environment of Serializer-Deserializer (SerDes) diversifies, the complexity of the system increases. In particular, Analog Front End (AFE) of SerDes receiver (Rx) is difficult to implement as a model. It is difficult to simulate all combinations of control settings. In addition, a complex adaptation algorithm is required to find the optimal value of AFE. Real circuit-like implementations for adaptation lead to long simulation period. These make accurate modeling difficult.
In this study, we evaluated a method of estimating the BER using the signal of the Rx input pad with the convolutional neural networks (CNN). We showed that the proposed model can estimate BER accurately. Our machine learning-based estimator can predict the performance of SerDes using signal and residual noise at the SerDes receiver. By training the data for each condition, it can be generalized and applied to different temperatures, process corners or SerDeses. The model in this study can be used to replace the existing IBIS-AMI model, and can also be used as an auxiliary estimator.