Using Transfer Learning on Deep Learning Networks to Predict Channel Operating Margin

The increasing demand for data-intense applications such as artificial intelligence, Internet of things (IoT), wearables, augmented reality and virtual reality has led to the need for the constant upgrade of data channels everywhere including in-server infrastructures and consumer electronics. These upgrades require carefully checking the signal integrity of connected (wired or wireless) devices. However, signal integrity analyses can be time consuming processes. Anything from solving 3-D electromagnetic models to performing time-domain analysis of channels can take several hours to a day or even longer. The advancements in CPU and GPU technology have helped but those advancements have been countered by the increased complexity of the components and channels being analyzed which are essential for the ever-increasing need for faster data rates and higher bandwidths. One avenue to decrease complex multi-channel analysis is to incorporate deep learning networks. After training, these networks would allow design space exploration of pre-layout designs and post-layout verification to be completed faster. In this paper, several deep learning networks are trained to look at various PAM4 eye diagrams and determine whether the channel will pass or fail the Channel Operating Margin (COM) without having to run the COM script and therefore allowing the designer to eliminate failing channels early in the design process. Results show that significant time savings can be achieved by using a retrained deep learning network over COM. This time savings translates into reduced design cycle development times.