HIGHER-LEVEL SYNTHESIS TO MACHINE LEARNING

Cong presented a full history of his group’s innovation including in-depth technical details, from their high-level synthesis from 2004 to today. In order to automate the chip design process, Cong and his group considered several methods. They worked systematically, identifying a method of higher-level synthesis (HLS) to use C and C++ as they create the chips.

Cong explained that design bottlenecks would occur frequently, so programmers added a snippet of code called a pragma where they occurred to tell HLS how to overcome the bottlenecks. However, it is time consuming to identify these bottlenecks, and not every software programmer has the knowledge or experience to add the correct pragmas. With deep learning, the system could instead make an educated guess and reduce the time needed to overcome the obstacles.

ChatGPT, a popular chatbot that leverages natural language processing, has largely entered national consciousness. Natural language processing works by synthesizing large amounts of data and producing educated guesses. While ChatGPT generates natural languages or code with ease, it does not have enough training data for integrated circuit design.

Another method, graph neural networks, may be the most appropriate for integrated circuit design, Cong said, largely because HLS is carried out on graphs — specifically the control and data flow graphs — and the GNN models are usually more compact and can be trained with less data. “We don’t need the system to do anything — just give an educated guess,” Cong said.

First, Cong and his team created a database to train the model to identify bottlenecks. Then, they built a predictive model using a single graph neural network encoder across all applications to train the system.

After replacing HLS with a deep learning model, Cong noted that his team was able to speed up the HLS optimization process from hours to milliseconds. The program, HARP, is intended to run with a Merlin compiler to help software programmers.

“Interacting with renowned faculty of Dr. Cong’s caliber helps us to stay informed and to consider possibilities for our own research.”

— Dr. Dinesh Bhatia,
Head and Professor of the Department of Electrical and Computer Engineering

Human intelligence first

“I spent the first part of my career learning to design hardware chips,” Cong said. “There are many architectures, such as systolic array designs, where human intelligence prevails. To do this, we need to have a good understanding of the optimal architecture, the structure design space and the property of optimization problem.”

Jonsson School leaders including Bhatia and Dr. Poras Balsara, vice dean of the Jonsson School and professor of electrical and computer engineering, invited Cong for the lecture.

Balsara has also worked on integrated circuits for his entire career, since the late 1980s. He observed that computer engineers have a fundamental understanding of the chip’s architecture that a software programmer may not possess and may be more adept at troubleshooting due to their familiarity with the technology.

While many questions remain, Cong’s talk illustrates numerous possibilities for technological advancement through deep learning that extends well beyond tools like ChatGPT.

“For those of us who have worked in the areas of electronic design automation and domain-specific architecture design, Dr. Cong has been an inspiration for us for the past 30 years,” said Dr. Benjamin Carrion-Schaefer, assistant dean for graduate student success and professor of electrical and computer engineering, who introduced Cong at the event. “Our colleagues and students especially appreciated visiting one-on-one with Dr. Cong.”

Discover More