Search for: All records

The input/ output modes are essential for high-speed signal integrity analysis and channel simulation. This work aims to develop a method for generating an IBIS-AMI model for USB 3.0 using measurement data. Instead of requiring a specially designed motherboard with test points for specific measurements, this method uses measurement data obtained from an assembled motherboard. The only available data for measurement in this case is the output voltage waveform from the USB 3.0 port on the motherboard. To address this, a novel approach is proposed to extract all the required parameters for the IBIS-AMI model from a single available measurement using a neural network. The neural network is trained with a set of IBIS-AMI models, each containing parameters with varying values, and a series of voltage waveforms generated from channel simulations with these IBIS-AMI models. Once trained, the neural network can generate the IBIS-AMI model using just one measured output voltage waveform. This constructed model has no limitations related to the output channel and can be applied to different output channels for analysis, making it a versatile tool for high-speed signal integrity evaluation. 

Wire bonding as a metallic interconnection is widely used to transmit high-speed signals and supply power within the integrated circuit (IC) packages. However, bonding wires also effectively radiate power noise and the harmonics of the output signals, causing electromagnetic interference and radio frequency interference issues. In this study, a current-loop model using a transfer admittance matrix for estimating the equivalent radiation sources of an IC/package featuring bonding wires is proposed. Based on the proposed modeling method, a novel reinforcement learning algorithm is applied to optimize the configurations of signal, power, and ground bonding wires, mitigating the radiation from the IC/package. The proposed modeling method is validated experimentally by a self-designed IC with an inverter-type buffer based on a complementary metal–oxide–semiconductor 0.18-μm process, and a radio frequency victim antenna built on the same printed circuit board. From 720 to 900 MHz, the maximum difference between the proposed modeling method and the measurement results is only 2.3 dB. In addition, full-wave simulation is performed to evaluate the optimization results of the reinforcement learning algorithm, showing radiation mitigation of over 7 dB compared to the randomly selected bonding-wire configurations.