Keynote: Cost-Efficient Reliability for Edge-AI Chips

Very recently, Artificial Intelligence started undergoing a remarkable transformation by moving closer to the source of data, thus establishing the Edge AI concept. This trend sets new reliability requirements for the related hardware chips used for safety- and mission-critical applications. The key research and engineering challenges stem from the limited computing and energy resources of the edge devices. Furthermore, the compute-efficiency and the cost of the reliability of the Edge-AI chips are becoming enabling factors for their way to the market. The talk discusses techniques for soft-error and lifetime reliability assessment and enhancement for Deep Learning accelerators. It advocates the role of approximate computing and looks into specifics of the systolic-array-, data-flow-based and industry-grade accelerator architectures for ASICs and FPGAs.