Abstract

In this paper, high-speed channel simulators using neural language models are proposed. Given the input sequence of geometry design parameters of differential channels, the proposed channel simulator predicts SI characteristic sequences such as insertion loss (IL) and far-end crosstalk (FEXT). Sequence-to-sequence (seq2seq) networks using a recurrent neural network (RNN) and a long short-term memory (LSTM) are utilized for the estimator. Moreover, a transformer network which is a recent neural engine of large language models (LLMs) is introduced for the first time. Compared to seq2seq networks, the transformer network-based simulator can achieve shorter computing time due to its parallel computation called an attention. The accuracy and training time of seq2seq and transformer networks are validated and compared. As a result, all the proposed simulators show ~1% error rates for both the IL and FEXT. However, for the training time, the transformer network achieves 75%-83% reduction compared to seq2seq networks.

Department(s)

Electrical and Computer Engineering

Comments

National Science Foundation, Grant IIP-1916535

Keywords and Phrases

High-speed channel; Neural language model; Signal integrity; Transformer network

International Standard Serial Number (ISSN)

2158-1118; 1077-4076

Document Type

Article - Conference proceedings

Document Version

Citation

File Type

text

Language(s)

English

Rights

© 2024 Institute of Electrical and Electronics Engineers, All rights reserved.

Publication Date

01 Jan 2024

Share

 
COinS