Testing the Divergence Stack Memory on GPGPUs: A Modular in-Field Test Strategy

General Purpose Graphic Processing Units (GPGPUs) are becoming a promising solution in safety-critical applications, e.g., in the automotive domain. In these applications, reliability and functional safety are relevant factors in the selection of devices to build the systems. Nowadays, many challenges are impacting the implementation of high-performance devices, such as GPGPUs. Moreover, there is the need for effective fault detection solutions to guarantee the correct in-field operation of a GPGPU, such as in the branch management unit, which is one of the most critical modules in this parallel architecture. Faults affecting this structure can heavily corrupt or even collapse the execution of an application on the GPGPU. In this work, we propose a non-invasive Software-Based Self-Test (SBST) solution to detect faults affecting the memory in the branch management unit of a GPGPU. We propose a scalar and modular mechanism to develop the test program as a combination of software functions. The FlexGripPlus model was employed to evaluate the proposed strategies experimentally. Results show that the proposed strategies are effective to test the target structure and detect up to 98% of permanent faults. General Purpose Graphic Processing Units (GPGPUs) are becoming a promising solution in safety-critical applications, e.g., in the automotive domain. In these applications, reliability and functional safety are relevant factors in the selection of devices to build the systems. Nowadays, many challenges are impacting the implementation of high-performance devices, such as GPGPUs. Moreover, there is the need for effective fault detection solutions to guarantee the correct in-field operation of a GPGPU, such as in the branch management unit, which is one of the most critical modules in this parallel architecture. Faults affecting this structure can heavily corrupt or even collapse the execution of an application on the GPGPU. In this work, we propose a non-invasive Software-Based Self-Test (SBST) solution to detect faults affecting the memory in the branch management unit of a GPGPU. We propose a scalar and modular mechanism to develop the test program as a combination of software functions. The FlexGripPlus model was employed to evaluate the proposed strategies experimentally. Results show that the proposed strategies are effective to test the target structure and detect up to 98% of permanent faults.