Machine learning approach for high speed link modeling and IBIS-AMI model generation

Wang, Xinying

Permalink

https://hdl.handle.net/2142/109343 Copy

Description

Title

Machine learning approach for high speed link modeling and IBIS-AMI model generation

Author(s)

Wang, Xinying

Issue Date

2020-10-07

Director of Research (if dissertation) or Advisor (if thesis)

Schutt-Aine, Jose E

Doctoral Committee Chair(s)

Schutt-Aine, Jose E

Committee Member(s)

• Cangellaris, Andreas C
• Chen, Deming
• Rosenbaum, Elyse

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

Machine Learning, Macro-modeling, High-speed Link, IBIS-AMI

Abstract

The high-speed link system is one of the major components in the networking infrastructure. Developing a high-performance behavioral model for such a system is crucial but challenging, especially when taking nonlinearity into account. This work reports modeling the high-speed link (HSL) system using machine learning and implementing the model into IBIS-AMI, an industrial standard for SerDes simulation and verification. We started with developing a Volterra series model by extracting the Volterra kernels using feed forward neural networks. We proposed a monomial power series neural network (MPSNN) which can extract Volterra kernels that relate to nonlinearity up to the third order. We developed an analytical mapping from neural network weights to Volterra kernels. The analytical mapping allows accurate time domain signal reconstruction with extracted Volterra kernels. We applied the MPSNN to model pulse amplitude modulation 4 level (PAM-4) and non-return-to-zero (NRZ) system. Volterra kernels up to the third order can be accurately identified. The curse of dimensionality associated with Volterra series impedes the practical applications of the Volterra series. The number of Volterra kernels increases exponentially with the increase in memory length and the nonlinearity order. The large number of Volterra kernels consume a vast amount of computational power during signal reconstruction. To address this challenge, we proposed a Laguerre-Volterra feed forward neural network (LVFFN). The input time-series signal is orthogonalized, in other words, Laguerre-expanded, before it is feed to the neural network. The dimension of the input signal is significantly reduced, which results in many fewer neurons in the hidden layer. We modeled the PAM-4 and NRZ system with LVFFN. The resulted model has the number of parameters that are up to six orders of magnitudes less than the Volterra series. We could also model just the receiver instead of the whole system to add more flexibility to the model in practical applications. The LVFFN model greatly addressed the curse of dimensionality associated with Volterra series. Then the next question addresses how are we going to use it. Since the machine learning based model is not a standardized model, it is difficult to be co-simulated with models generated by other approaches. To circumvent the challenges in model transportability and interoperability, we implemented the LVFFN model into the IBIS-AMI model, an industrial standard model that is compatible with most of the circuit simulators. We could simulate the LVFFN IBIS-AMI model in Keysight ADS and conduct the eye-diagram analysis. IBIS-AMI model generation is not trivial. It requires cross-disciplinary knowledge in signal integrity, HSL circuit, and software engineering. To facilitate the process of IBIS-AMI model generation, we developed a software, ezAMI, that can generate the IBIS-AMI model by clicks. The software is developed using Qt/C++ and is an open-source software. The software architecture and tutorial are introduced in this dissertation as well.

Graduation Semester

2020-12

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/109343

Copyright and License Information

Copyright 2020 Xinying Wang

Owning Collections