RunSAFER: A Novel Runtime Fault Detection Approach for Systolic Array Accelerators

In this study, we introduce a new runtime fault detection technique for systolic array accelerators oriented to neural network applications. The method exploits the functional path of the systolic array to compute and process checksum values during the execution of the current application instructions flow and integrates self-testing capabilities within the systolic array Instruction Set Architecture. The proposed technique does not require additional hardware self-testing modules, and the test pattern penalty is limited to 3 clock cycles independent of the size of the systolic array. Experimental analysis performed with fault injection campaigns demonstrates full fault detection capabilities of stuck-at faults with an average computing overhead 4 times lower than state-of-the-art solutions. Additionally, our approach exhibits diminished hardware overhead in contrast to conventional techniques.