Oak Ridge National Lab (ORNL): A Spiking Neural Network Architecture for Ultra-low Power and Ultra-low Latency Computing

The U.S. Department of Energy (DOE) has called on the national labs to strengthen their microelectronics capabilities. In response, Oak Ridge National Lab (ORNL) has formed an internal microelectronics initiative with a focus on AI for scientific applications and energy-efficient computing. Researchers in the Architectures and Performance Group at ORNL have developed a digital neuromorphic architecture specifically designed for scientific applications that require ultra-low latency. With the help of the Siemens EDA tools, including Catapult HLS, the neuromorphic accelerator is being adapted from an FPGA prototype to a more capable and lower-power ASIC implementation. In collaboration with the Sensors and Electronics Group, the neuromorphic processor is also being re-target towards another import DoE mission of nuclear safeguards. This effort is exercising Siemens analog and mixed-signal tooling to integrate the neuromorphic processor with the sensing elements more closely by exploring efficient analog encoding neurons and event-driven gamma-ray detection. This presentation will give a brief overview of nuclear safeguards and other scientific applications at ORNL. Then, we will present the work done so far to develop the ASIC neuromorphic processor, the progress and plans for the mixed-signal processing and sensor subsystem, and finally, the dataset and algorithmic details we are leveraging to demonstrate the project’s capabilities.