1052: AI Based Media Functional Safety and Reliability Verification in Safety-Critical Autonomous Systems

Thanks to the advancements in processor compute, connectivity, and memory technology, the demand for video processing silicon is on the rise over the past few years. In addition to the traditional video entertainment applications, the explosive growth in AI technology enabled several machine vision applications as well. Today’s Systems on Chips (SoCs) catering to these various multimedia applications are a complex chain of hardware with significant memory requirements. The importance of the functional Safety and reliability verification in the design is very critical and it has the possibility for catastrophic issues for the end users if it’s not verified properly. The impact of soft errors in the silicon especially during the memory transactions to the end user experience and to the inference accuracy of the machine vision algorithm is largely an unexplored area. Our paper tries to bridge this gap by outlining a novel methodology and a unique verification framework that provides the ability to inject soft errors and study the impact of those errors on both video/image quality as well as on the inference accuracy. Using this verification framework, we studied the impact of soft errors in memory for the JPEG encoder hardware and expanded the study to include video encode hardware, where the encoded video streams will be used for both human consumption as well as for machine vision. For human consumption, the framework computes both objective video quality using standard metrics as well as subjective quality metrics using VMAF (Video Multi-Method Assessment Fusion). For machine vision, the framework executes AI workloads including object detection, tracking and classification and compute various metrics such as mean average precision, multi-object tracker accuracy as well as classification accuracy. The proposed verification methodology uses the power of SystemC and OpenVINO which provides a novelty in functional safety and reliability Verification.