New Models Make Circuit Design More Efficient

Posted on Tuesday, October 12th, 2021

Image of a vehicle rearview camera with a hand on a steering wheel.
Field Programmable Gate Arrays (FPGAs) are circuits that can be configured after manufacturing. For example, if regulations change for rear-view cameras, this tech allows for reprogramming rather than recall.

Machine Learning models streamline circuit design, with applications ranging from electronics to aerospace.

From consumer electronics to aerospace, “Field Programmable Gate Arrays” have impressive applications. These circuits have important design considerations, including where to place the required electrical components within a limited amount of space—"placement”—as well as the design of the connection wires—"routability.” For example, have you ever wondered how all the necessary components and connections fit inside your vehicle’s rear-view camera?

Using Machines to Make Predictions

University of Guelph researchers, including Dr. Shawki Areibi (School of Engineering) as well as Dr. Gary Grewal and Master’s student Timothy Martin (School of Computer Science), are using machine learning models to predict placement and routability more effectively. Previously, these researchers used deep learning—a type of machine learning modelled after the human brain—to develop a model to determine successful circuit routing and placement without having to invest a lot of computer power. Their model produced accurate predictions within milliseconds. However, Deep Learning models require significant effort on the front end, including access to large datasets and time spent training the computers.

In their work—recently nominated for best paper at the 3rd ACM/IEEE Workshop on Machine Learning for CAD a Machine Learning on CAD Workshop—the team seeks to show that simpler Machine Learning models can be as effective as Deep Learning for predicting the routability of a placement, saving time and effort.

Modelling and Experimenting

The team developed several Machine Learning models and applied an experimental approach to determine the features best used to train the model (for example, estimated wirelength). They then investigated five Machine Learning algorithms and three techniques that combined multiple models.

Improved Efficiency

The researchers found that when Machine Learning models are trained with the right features, they achieve higher accuracy when compared to the Deep Learning method. Not only that, but they require a fraction of the training data and Central Processing time, meaning they are much more efficient.

“Machine Learning models can improve the accuracy and efficiency of routability prediction models, important in the Computer-Aided Design Flow of Field Programmable Gate Arrays (FPGAs),” says Areibi. “In future, we plan to explore how these probes and prediction tools can be used to improve both the quality of the result and to reduce the compile time of Field Programmable Gate Arrays.”

Headshot of Dr. Shawki Areibi

Shawki Areibi is a Professor in the School of Engineering.

Headshot of Dr. Gary Grewal

Gary Grewal is an Associate Professor in the School of Computer Science.

This work was supported by the Natural Sciences and Engineering Research Council of Canada.

Martin T, Areibi S, Grewal G. Effective Machine-Learning Models for Predicting Routability During FPGA Placement. In 2021 ACM/IEEE 3rd Workshop on Machine Learning for CAD (MLCAD) (pp. 1-6). IEEE.

Read more about Dr. Areibi and Dr. Grewal's award-winning research.

News Archive