Achieving diverse redundancy for GPU Kernels

Autonomous driving requires high-performance computing devices including general-purpose CPUs as well specific accelerators, with GPUs having a key role due to their flexibility. Safety-critical microcontrollers have achieved ASIL-D compliance by implementing diverse redundancy lockstep execution on-chip. However, GPU does not provide natively, thus failing reach ASIL-D, which could only be reached fully redundant lockstepped (2 GPUs) or pairing another accelerator. both options may infeasible procurement costs, and additional power, space reliability costs accomodate two devices. In this work, we present variety of solutions enable using one taking advantage the already internal GPUs. We lowly-intrusive hardware software-only solution, latter evaluated directly on real platform. case kernel require tailoring some parameters. With that objective, also propose an algorithm performs such automatically guarantee Overall, our allow achieving single either Commercial off-the-shelf GPU, in more efficient manner introducing minor changes design.