About the Functional In-Field Self-Testing of AI Accelerators

<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>Work supported by the National Resilience and Recovery Plan (PNRR) through the National Center for HPC, Big Data and Quantum Computing.Deep Learning Accelerators (DLA) are pervasive hardware units in modern applications, including safety critical systems such as Automotive, Aerospace, Robotics, and health monitoring systems. Therefore, it is crucial to guarantee their reliability during the mission operation of any of these applications, as a failure can produce catastrophic results (e.g., loss of human lives). In fact, modern semiconductor technologies used to implement DLAs can be affected by faults due to several phenomena, such as aging, process variation, or manufacturing defects. Periodic testing and functional testing strategies have demonstrated their utility and effectiveness on DLA accelerators in spotting faults arising during the in-field operation of the system. Nonetheless, these approaches can have significant testing times and memory footprint overheads, making them hard to apply during the online operation of large-size accelerators. This work proposes an effective strategy for generating efficient functional test patterns for computational units in DLA accelerators by reducing the required testing time (43×) and memory footprints (3.3×) compared with literature solutions. Moreover, our strategy provides diagnostic capabilities for identifying defective units (e.g., multipliers).