Modular Neural Network-Based Models of High-Speed Link Transceivers

Abstract

In this article, we address the nonlinear behavioral modeling of transceivers using feedforward neural networks (FNNs) such that each modular block functions independently in a high-speed link (HSL) simulation. In the proposed technique, the modular transceiver models are represented in the form of kernel matrices, in which the values are determined through FNN training. By feeding the FNN models with information on voltages and protocols, the nonlinear time-domain HSL analysis is transferred to simple matrix multiplications, which allows significant simulation speedup while preserving good accuracy. Compared to the conventional modeling standards, the I/O buffer information specification (IBIS) or IBIS-AMI models, the generation of FNN models requires minimal effort, thereby permitting wider access to the technique. Furthermore, we demonstrate that transceiver modeling with FNN is highly robust and flexible in terms of feature expansion. With minor adjustments in the protocols, advanced settings, such as equalization and differential signaling, can be easily included in the trained FNN models.