While artificial intelligence has advanced rapidly in recent years to become a new frontier of technology, its capabilities wouldn’t be possible without the underlying computer hardware.

The heart of AI hardware includes custom silicon chips known as accelerators.

Traditionally, designing AI accelerator chips has required human experts with knowledge in machine learning, computer architecture and very large-scale integration, or VLSI. Jeff Zhang, an assistant professor of electrical and computer engineering in the Ira A. Fulton Schools of Engineering at Arizona State University, set out to reduce the labor-intensive AI hardware development process with a team led by Antonino Tumeo, a computer scientist at the Pacific Northwest National Laboratory.

Zhang, a faculty member in the School of Electrical, Computer and Energy Engineering, part of the Fulton Schools, and his collaborators’ efforts resulted in the software-defined architecture, or SODA, Synthesizer, a novel electronic design automation tool that automatically completes the AI accelerator design process. The resulting paper, published in the Institute of Electrical and Electronics Engineers scientific journal IEEE Micro, received the journal’s 2022 Best Paper Award.

“This award recognizes our team’s vision and efforts in future hardware design,” Zhang says.

The SODA Synthesizer can generate specialized machine learning accelerator chips from a high-level description of the desired specifications.

“Such tools will close the gap of lab-to-fab solutions for the future of semiconductor technology,” he says.

SODA builds upon a set of open-source infrastructures, including the Multi-Level Intermediate Representation, or MLIR, platform, and a high-level synthesis tool called Bambu. It can leverage compiler optimization techniques to explore the hardware design space for better power, area and performance.

Zhang and his students plan to build on the research from the SODA project, to streamline the computer hardware design process.

Those interested in trying SODA can download the necessary files and information from GitHub.