PNNL

Computations play a huge role in solving complex scientific problems. The machines that run these computations—including data analysis, machine learning, and scientific computing algorithms—need to provide high performance at low energy costs. Accelerators can boost computational performance, but many domain scientists do not have the know-how to design their own accelerators to meet their project’s needs.

Researchers at Pacific Northwest National Laboratory (PNNL) created the software-defined accelerator (SODA) toolkit to empower researchers to design their own custom accelerators. Their paper, published in IEEE Micro, was awarded a 2022 Best Paper Award by the IEEE Computer Society Publications Board.

“Designing custom accelerators used to be difficult and time consuming—most domain scientists aren’t familiar enough with the hardware description languages needed to create them,” said Antonino Tumeo, a computer scientist and co-author of the paper. “SODA allows users to explore different accelerator designs without requiring any hardware design expertise.”

SODA’s compiler-based front end interfaces with high-level programming frameworks and applies high-level optimizations to generate a code for its back-end framework, Bambu. The high-level synthesis back end then compiles the final design of the accelerator. Users can select different command line options to explore different accelerator designs.

While this framework is useful across all scientific domains, the researchers see a distinct need for custom accelerators combined with artificial intelligence (AI), particularly deep neural networks (DNNs) and autonomous experimentation.

“After a DNN model is generated with Python—Pytorch or Tensorflow—and translated into our compiler framework, the SODA toolchain takes place, enabling domain scientists to generate hardware specific for their machine learning models,” said Nicolas Bohm Agostini, lead author of the paper.

The researchers are now exploring ways to adapt accelerators in an automated way. As part of the Adaptive Tunability for Synthesis and Control via Autonomous Learning on Edge (AT SCALE) initiative, the researchers are exploring the end-to-end codesign of highly specialized and low-latency hardware accelerators needed for reasoning and control. These would be used in closed-loop precision synthesis platforms, such as those in autonomous electron microscopes.

“Without custom accelerators, progress in autonomous experimentation will be limited,” said Tumeo. “The open source SODA toolkit removes barriers for researchers to design their own chips.”

This research was supported by Pacific Northwest National Laboratory’s Data-Model Convergence Initiative. Additional PNNL authors on the paper are Serena Curzel, Ankur Limaye, Vinay Amatya, Marco Minutoli, Vito Giovanni Castellana, and Joseph Manzano. Jeff (Jun) Zhang, David Brooks, and Gu-Yeon Wei of Harvard University, and Cheng Tan from Microsoft also contributed to this work. PNNL is advancing work in AI technology through its newly launched Center for AI.