Software-Only Triple Diverse Redundancy on GPUs for Autonomous Driving Platforms

Autonomous driving (AD) imposes the need for safe computations in high-performance computing (HPC) components such as GPUs, thus with capabilities to detect and recover from errors since a safe state may not exist anymore. This can be achieved with Triple Modular Redundancy (TMR) for computation components. Furthermore, error detection capabilities need to provide some form of diversity to avoid the case where a single fault leads all redundant executions lead to the same error, which would go undetected. In our past work, we assessed GPUs against dual modular redundancy (DMR) with diversity, showing their potential and limitations to provide diverse redundancy building on reset and restart for recovery. However, such recovery scheme may be too slow for some applications. This paper proposes a software-only solution to deliver diverse TMR on commercial off-the-shelf (COTS) GPUs. Our work details how staggered execution can be achieved and assesses the performance of TMR on COTS GPUs. Moreover, we identify those elements where diversity cannot be guaranteed and provide some discussion comparing the case of DMR and TMR for those elements.