On the in-field test of the GPGPU scheduler memory

GPGPUs have been increasingly successful in the past years in many application domains, due to their high parallel processing capabilities and energy performance. More recently, they started to be used in areas (such as automotive) where safety is also an important parameter. However, their architectural complexity and advanced technology level create challenges when matching the required reliability targets. This requires devising solutions to perform in-field test, thus allowing the systematic detection of possible permanent faults. These faults are caused by aging or external factors that affect the application execution and potentially generate critical misbehaviors. Moreover, effective infield test techniques oriented to verify the integrity of GPGPU modules during in-field operation are still missed. In this work, we propose a method to generate self-test procedures able to detect all static faults affecting the scheduler memory existing in each streaming multiprocessor (SM) of a GPGPU. NVIDIA CUDA-C is selected as high-level programing language. The experimental results are obtained employing the NVIDIA Nsight Debugger on a NVIDIA-GEFORCE GTX GPU and a memory fault simulator