C-SMART: A preprocessor for neural network performance and reliability under radiation

• Published in: Microelectronics and reliability Vol. 173; p. 115859
• Main Authors: Justus Rajappa, Anuj, Reiter, Philippe, Rech, Paolo, Mercelis, Siegfried, Famaey, Jeroen
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.10.2025
• Subjects: Bit-flips; Machine learning; Neural network; Preprocessor; Radiation; Reliability; Neural network; Preprocessor; Bit-flips; Reliability; Machine learning; Radiation
• ISSN: 0026-2714
• DOI: 10.1016/j.microrel.2025.115859

Edge AI brings the benefits of AI, such as neural networks for computer vision analysis, to low-power edge computing platforms. However, application and resource constraints leading to inadequate protection can make edge devices vulnerable to environmental factors, such as cosmic rays that continually shower on Earth. These factors can cause bit-flips that affect the reliability of the neural network inferences computed using these edge devices. To address this issue, we developed the Conditional-SMART (C-SMART) preprocessor designed to answer the question ‘When to use SMART?’, for obtaining both reliability and performance benefits. SMART is a reliability improvement technique introduced in our previous work, which involves skipping the multiply–accumulate operations performed on the zero-valued inputs to the layers of the neural network. We demonstrated C-SMART with a commercial bare-metal system containing an ARM microprocessor by exposing the system to real-world, atmospheric-like neutron radiation using the ChipIr facility in Oxfordshire, UK. We also conducted timing and energy measurements for performance analysis. Our experiments with C-SMART for inference with a neural network revealed a reliability boost against soft errors by more than 26% while improving performance by more than 35%. We foresee these benefits in various COTS devices by integrating C-SMART with compilers and neural network generators. •A compiler-level preprocessor for AI efficiency and reliability against radiation.•Reliability improved by 26% and performance improved by 35%.•Tested using atmospheric-like neutrons for terrestrial applications.