Toward a generic and flexible architecture for AI hardware
Résumé
Environmental conditions, low power, high connec- tivity and heterogeneous nature of internet of things (IoT) make it very vulnerable to cyberattacks and failures. In addition, security and reliability can influence each other, especially, countermeasures can increase one notion at the cost of the other. With the development of IoT in critical fields such as energy, transportation or healthcare, safety has become a bigger concern. It is therefore paramount to evaluate, not only security and reliability separately, but also to measure their interactions. In this work, we propose to investigate the challenge of assessing security measures’ reliability, and reliability measures’ security. Our goal is to develop a methodology that gives quan- titative indicators on the modifications of security and reliability induced by a countermeasure. With this evaluation, architects will be able to compare several implementations with security and reliability in mind.This paper provides an overview of AI algorithms and hardware accelerators. We examine AI algorithms focusing on their computational needs and areas of application. Fur- thermore, we analyze various AI accelerators, highlighting their flexibility, and suitability for different AI workloads.