Massively parallel processors provide high computing performance by increasing the number of concurrent execution units. Moreover, the transistor technology evolves to higher density, higher frequency and lower voltage. The combination of these factors increases significantly the probability of hardware failures. In this paper, we present a methodology to locate and mitigate hardware failures of NVidia GPUs. Results show that intermittent errors can be precisely localized and have a limited impact to a well defined architecture tile. Therefore, we propose, and demonstrate on a software prototype, a rescheduling strategy to quarantine the defective hardware and ensure correct execution. Our approach significantly improve the GPU fault-tolerance capability and GPU’s lifespan, at a reasonable overhead.