Model Establishment and Deep-Neural-Network Solutions for Signal Integrity in DDR Systems

• Main Authors: Kang-Yun Yang, 楊岡昀
• Other Authors: Ruey-Beei Wu
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/kcmy8u

博士 === 國立臺灣大學 === 電信工程學研究所 === 106 === The double data rate (DDR) memory system has been widely used in many electronic devices. As its data rate has reached the Gigabit range and many signal integrity (SI) problems arise in the signal lines. In this thesis, SI problems of three kinds of topologies are discussed and its equivalent models are established. First, single-ended lossy transmission line is discussed. Single-ended signal lines such as data lines (DQ) can be simplified into a lossy transmission line. The main SI issue should be the conductor loss and dielectric loss. Second, tree topology is studied. Finally is the fly-by topology. The research results reported in this thesis are organized as follows. First, the fundamental theory of the lossy transmission line is described. After that, some useful formulas are given for the approximations of both conductive and dielectric losses in PCB-scale transmission lines. The simplified transfer function and thus the corresponding impulse response will be analytically derived as well. Next, by using rational function fitting, the general equivalent circuit can be built for any tx-line parameters. Finally, eye diagram is calculated and the relation to tx-line parameters is discussed. Next, the research background and the mechanism of the ringing effects on branch line structure are presented. Then, an RLC model is built to characterize the ringing phenomenon. Finally, the built model is used to discuss the worst case eye height of the branch line and a design guide is given to optimize the signal integrity. The crosstalk noise of interconnects is also taken into account. First, a theoretical method to build an equivalent model similar to the single-ended line is proposed as a generalized version. Next, a fast methodology that predicts the worst-case crosstalk interconnect using coupling coefficient from inductance and capacitance matrices is proposed. Finally, a practical case is demonstrated and some useful design guides are provided to the cooperative company. For cases that is too complicated to derive a physical equivalent model, an automatic design flows using Matlab to access the optimal design of interconnects is proposed. This flow uses Matlab to write Hspice scripts and run the transient simulation automatically. After some training data are calculated, artificial neural network (ANN) is proposed to find a good fitting function of the optimization design. After a general optimization flow was developed, a novel dual sided fly-by topology, which is derived from fly-by, is proposed. To facilitate the design, the research utilizes the Matlab automatic flow to acquire optimal design parameters. Conclusively, the electromagnetic bandgap is employed to explain the physical mechanism. The method to minimize the bandgap so as to achieve higher eye height is provided as well.