High-Speed Receiver Transient Modeling with Generative Adversarial Networks

Data-intensive applications such as artificial intelligence and graph processing are becoming commonplace, requiring high-speed IO to enable the deployment of these critical applications. To accommodate the increasing data requirements Serializer/Deserializer (SerDes) receivers have become increasingly complex, with different equalization schemes to mitigate channel impairments. It has become increasingly important to model this receiver as they are performance-critical. This paper shows an approach to modeling the transient of a high-speed receiver with fixed and varying equalization through generative networks. The method considers the receiver as a black box, with its inputs and outputs as two different domains, framing the problem as a domain translation task. The proposed approach uses an intermediate representation of the time series to model the receiver successfully. We demonstrate that the proposed method is invariant to the input waveform, receiver configuration, and channel. In a fixed equalization setting, the proposed approach has a root-mean-squared error of 0.016 in a [0,1] range and an error of 0.054 in the same range for a variable redriver. The approach can predict a batched set of results under 250ms, faster than an equivalent spice model for the same time steps.