Software-only diverse redundancy on GPUs for autonomous driving platforms

Autonomous driving (AD) builds upon high-performance computing platforms including (1) general purpose CPUs as well as (2) specific accelerators, being GPUs one of the main representatives. Microcontrollers have reached ASIL-D compliance by implementing diverse redundancy with lockstep execution. However, ASIL-D compliant GPUs rely on either fully redundant lockstep GPUs (i.e. 2 GPUs), which doubles hardware costs, or fully redundant systems with a GPU and another accelerator, which virtually doubles design and validation/verification (V&V) costs. In this paper we analyze the degree of diversity achieved when implementing redundancy on a single GPU, showing that diverse redundancy is not achieved in many cases, and propose software strategies that guarantee achieving diverse redundancy for any kernel on systems using commercial off-the-shelf (COTS) GPUs, thus showing how to achieve ASIL-D compliance on a single COTS GPU in controlled scenarios.This work has been partially supported by the Spanish Ministry of Economy and Competitiveness (MINECO) under grant TIN2015-65316-P and the HiPEAC Network of Excellence. Jaume Abella has been partially supported by the MINECO under Ramon y Cajal postdoctoral fellowship number RYC2013-14717Peer ReviewedPostprint (author's final draft