Aging and Soft Error Resilience in Reconfigurable CNN Accelerators Employing a Multi-Purpose On-Chip Sensor

Convolutional Neural Networks (CNNs) have found a wide range of applications beyond conventional image processing, penetrating safety-critical domains such as automotive, medical, and space applications. This indicates that the demand for adaptable/reconfigurable hardware combined with reliable operation becomes imperative. In this paper, we propose the use of a multipurpose on-chip sensor capable of detecting aging and soft errors in order to enable CNN reconfigurability during the runtime. In this scenario, the sensor plays a pivotal role by providing in-field real-time intelligence on the health of the CNN hardware accelerator. Moreover, we introduce the concept of a reconfigurable CNN accelerator based on shared layers methodology, which can operate in fault-tolerant (FT), high-performance (HP), and de-stress (DS) modes. This study further delves into an architectural modification to integrate the sensor in a reconfigurable CNN accelerator, enabling it to adjust dynamically to changing AI workload needs. The sensor working principle and integration with the CNN accelerator architecture are discussed. Experimental results based on a commercial FPGA implementation demonstrate that the integration of the on-chip sensor yields a negligible penalty in terms of the CNN accelerator’s area overhead, power consumption, and response latency. This study aims to establish a foundation for the development of fully reconfigurable AI processing systems.