A software-only approach to enable diverse redundancy on Intel GPUs for safety-related kernels

Abstract

Autonomous Driving (AD) systems rely on object detection and tracking algorithms that require processing high volumes of data at high frequency. High-performance graphics processing units (GPUs) have been shown to provide the required computing performance. AD also carries functional safety requirements such as diverse redundancy for critical software tasks like object detection. This implies that software must be executed redundantly (in a single GPU for efficiency reasons), and with some form of diversity so that a single fault does not cause the same error in both redundant executions. Unfortunately, high-performance GPUs lack explicit hardware means for diverse redundancy, and software-based solutions with limited guarantees have only been provided for NVIDIA GPUs. This paper presents a software-only solution to enable diverse redundancy on Intel GPUs achieving, for the first time, strong guarantees on the diversity provided. By smartly tailoring workload geometry and managing workload allocation to execution units with thread-level wrappers, we guarantee that redundant threads use physically diverse execution units, hence meeting diverse redundancy requirements with affordable performance overheads.