Achieving diverse redundancy for GPU Kernels

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