Athena Seminar Series: Edge AI acceleration: From efficient models to efficient circuits

In the landscape of Artificial Intelligence (AI), two complementary paradigms have emerged: cloud AI, which relies on centralized computing infrastructure, and edge AI, which performs intelligence directly on distributed devices at the network edge. Driven by the rapid growth of wearables, drones, and other Internet-of-Things (IoT) devices, edge AI is increasingly pervasive and transforming the way intelligent services are delivered in everyday life. This talk will focus on two key components of the edge AI design stack that critically influence the performance and energy efficiency of AI acceleration. First, we will discuss hardware-aware Neural Architecture Search (NAS), an automated framework for jointly optimizing neural network architectures and computing architectures for target workloads and deployment constraints. We will then explore recent advances in digital Compute-In-Memory (CIM) and Compute-In-Logic (CIL) circuits, highlighting how architectural and circuit-level innovations can substantially improve the efficiency of AI computation. Finally, we will share how these approaches could potentially provide a practical pathway toward energy-efficient AI systems for next-generation edge platforms.