Public Lecture #3

"From Storage to Learning: New Memory Hierarchies for the Edge AI Revolution"

Prof. David ATIENZA ALONSO (École Polytechnique Fédérale de Lausanne (EPFL))

Abstract

Edge computing is becoming essential as Generative AI drives the need for systems capable not only of inference but also of local training and adaptation. Enabling on device learning at ultra low power requires a fundamental redesign of memory hierarchies, which remain the dominant bottleneck in modern edge AI platforms. Deep Learning and emerging GenAI workloads demand architectures that minimize data movement, efficiently support gradient updates, and exploit heterogeneous memory structures implemented in the latest semiconductor technology nodes, where density, leakage, variability, and interconnect constraints are critical.

This keynote introduces a vision for memory-driven edge AI architectures tailored for advanced nodes, integrating Processing-in-Memory (PIM), Near-Memory Computing (NMC), and reconfigurable scratchpads with embedded low-precision learning engines. These architectures are co designed with next generation memory technologies to balance bandwidth, energy efficiency, and reliability while remaining scalable. All these concepts will be illustrated within the open-source X-HEEP platform, which has been extended with memory-centric modules that enable lightweight training for diverse applications and real-time personalization of medical systems, thereby enabling a new class of adaptive and sustainable edge AI systems.