An apparatus to facilitate runtime fault detection, fault location, and circuit recovery in an accelerator device is disclosed. In one implementation, the accelerator device comprises a sensor network comprising a plurality of sensors; a secure device manager (SDM); and a sensor aggregator communicably coupled to the sensor network and the SDM. In one implementation, the sensor aggregator can receive sensor data from the sensor network; analyze the sensor data to detect a fault condition; determine a spatial location of the fault condition based on the sensor data; and generate an event for the SDM to cause the SDM to mitigate the fault condition.