An on-line testing technique for the scheduler memory of a GPGPU

The highly parallel processing capabilities and reduced power performance of General Purpose Graphics Processing Units (GPGPUs) have been crucial factors for their massive use in multiple fields, such as multimedia and high-performance computing applications. Nowadays, more demanding areas, such as automotive, employ GPGPU devices where safety and reliability are mandatory design constraints. Nevertheless, the structural complexity, the transistor density, and the implementation in the latest silicon technologies introduce challenges to match safety and reliability requirements. In these technologies, wear-out and aging are factors that may significantly increase the occurrence of permanent faults during the lifetime operation. Moreover, these faults may generate unacceptable misbehaviors during the execution of an application. These constraints require devising new methods for in-field fault detection, thus verifying the integrity and correct behavior of the device during its whole operational life. This work proposes a technique to generate functional self-test programs targeting the detection of permanent static faults in the memory of the warp scheduler of a GPGPU. The proposed technique can translate fault primitives, which represent the effect of faults in a memory cell, into self-test functions and programs composed of a sequence of operations to excite the fault in the memory and to propagate its effects to a visible location, thus detecting its presence. We focused on the memory in the warp scheduler because it represents a crucial module for the device operation. Furthermore, this memory is present in each Streaming Multiprocessor (SM) of a GPGPU. Some experimental results to validate the method have been gathered, resorting to the NVIDIA Visual Profiler and the Nsight Debugger using the NVIDIA-GEFORCE GTX GPU platform and a structural fault simulator. The CUDA programming environment was used to implement the test procedures.