International Conference on Design Automation Celebrates Aggie Engineers for Most Influential Paper of Past Decade
In 2016, a team of electrical and computer engineers at the University of California, Davis, set the stage for the efficient implementation of trained deep convolutional neural networks with a groundbreaking paper published in IEEE Xplore.
Titled “Design Space Exploration of FPGA-Based Deep Convolutional Neural Networks,” the article has now received the 10-year Retrospective Most Influential Paper Award from the Asia South Pacific Design Automation Conference, or ASP-DAC, one of the largest international conferences on silicon semiconductor research.
The award recognizes high-impact research that has made a significant contribution to the ASP-DAC community over the past decade. The UC Davis team, consisting of then-graduate students Mohammad Motademi and Philipp Gysel and Professors Soheil Ghiasi and Venkatesh Akella, presented their research during the 2016 IEEE/ACM ASP-DAC conference.
At a time when neural networks had only recently regained popularity after researchers struggled to advance the technology in the two preceding decades, the paper proposed a methodical solution to one of the biggest hurdles facing AI: execution speed.
The paper describes a methodology for designing a high-performance deep convolutional neural network on a field-programmable gate array, or FPGA — a kind of chip Ghiasi compares to a blank canvas. FPGAs come with logic blocks and memory units built in, enabling researchers to program the chips as they see fit, even allowing for flexibility for the chip’s function to change down the road.
The UC Davis team was not the first to use FPGAs for implementation of neural networks, but they were among the first to systematically demonstrate how effective they could be at powering AI technologies. The design method they developed leveraged the sources of parallelism in convolutional neural networks at multiple scales to achieve an approximately 100 percent increase in processing speed over the then-state-of-the-art design method.
At low to medium volume, FPGAs are also cheaper than a bespoke chip designed in-house. For that reason, FPGAs are now a critical piece of the AI industry, with Intel and AMD the two production leaders, according to Ghiasi.
“Our paper was at the forefront for accelerating image processing models,” Ghiasi said. “People have turned to it as a framework [for their own advancements in AI].”
The paper also marks the significant collaboration between graduate students Motademi and Gysel, as well as the start of their outstanding careers in industry.
Motademi was the first author on this FPGA-based research, which served as the foundation for Gysel’s paper on Ristretto, a hardware-aware optimization technique for convolutional neural networks, published in IEEE Transactions on Neural Networks and Learning Systems in 2018.
“[Our FPGA] work and Philipp’s outstanding work on Ristretto were among the early publications on hardware-aware optimization of modern deep neural networks, and the approaches used in these two works are now widely practiced in industry,” Motademi said.
Since graduation, Motademi has worked at NVIDIA, the world leader in AI computing technologies (to put the AI boom in perspective, NVIDIA was valued at 50 cents a share when Motademi’s paper was submitted to ASP-DAC in 2015; it is now hovering around $200 a share in 2026), while Gysel has worked at Qualcomm to advance machine learning algorithms.
“We’re honored that ‘Design Space Exploration of FPGA-Based Deep Convolutional Neural Networks’ has received the 10-year Most Influential Paper Award,” Venkatesh Akella said, one of the co-authors on the paper. “This recognition is especially meaningful as long-term validation from the community, and a proud moment for our [former] graduate students, Mohammad Motamedi and Phillip Gysel.”
