A System-level Abstraction and Service for Flourishing AI-powered Applications

Modern devices increasingly employ local AI accelerators such as GPUs, NPUs, and ASICs to meet growing demands for AI capabilities. However, the current ad-hoc and siloed approach to enabling AI poses two fundamental challenges. Developers have to handle unnecessary implementation details when integrating AI into individual applications. Moreover, the siloed approach hinders efficient resource management when running multiple AI-powered tasks simultaneously.