Memory System Design for AI/ML & ML/AI for Memory System Design: SRC AIHW Annual Review - 22.05.2023

Onur Mutlu Lectures
Onur Mutlu LecturesFeb 26, 2026

Why It Matters

Redesigning memory systems unlocks AI’s next performance leap while addressing security and power challenges, reshaping the semiconductor landscape.

Key Takeaways

  • AI workloads expose memory bandwidth bottlenecks
  • Processing‑in‑memory reduces data movement overhead
  • RowHammer remains a critical security concern
  • Intelligent controllers enable adaptive memory behavior
  • Cross‑layer co‑design drives energy‑efficient AI

Pulse Analysis

The explosive growth of AI and machine‑learning models has exposed the limitations of conventional memory hierarchies, where data shuttles between storage and compute units, inflating latency and power consumption. Researchers, led by Prof. Onur Mutlu, propose a memory‑centric paradigm that embeds compute capabilities directly within DRAM or emerging non‑volatile memories. By processing data where it resides, these processing‑in‑memory (PIM) solutions cut data movement by orders of magnitude, delivering faster inference and training while dramatically lowering energy per operation. This architectural shift also aligns with the industry’s push toward heterogeneous integration, where specialized accelerators and smart memory modules coexist on a single package.

Security considerations accompany this transformation. The RowHammer phenomenon—where repeated activations of a DRAM row cause bit flips in adjacent rows—has evolved from a curiosity into a systemic vulnerability. Mutlu’s work not only chronicles the retrospective analysis of RowHammer but also presents mitigation strategies such as adaptive refresh, error‑correcting codes, and intelligent memory controllers that dynamically adjust access patterns. These safeguards are crucial as PIM designs increase the density and proximity of active circuitry, potentially amplifying fault injection risks if left unchecked.

Looking ahead, the convergence of intelligent memory controllers, cross‑layer co‑design, and advanced materials promises a new generation of AI‑optimized systems. By integrating algorithmic awareness into the memory stack, hardware can prioritize critical data, prefetch intelligently, and allocate resources based on workload characteristics. This holistic approach not only boosts throughput but also extends the viability of Moore’s Law through architectural innovation rather than sheer transistor scaling. Companies that adopt memory‑centric designs will gain a competitive edge in delivering AI services with lower latency, higher security, and reduced operational costs.

Original Description

Title: Memory System Design for AI/ML & ML/AI for Memory System Design
Venue: SRC AIHW Annual Review / Task Kickoff
Presenter: Professor Onur Mutlu (https://people.inf.ethz.ch/omutlu/)
Date: May 22, 2023
Slides (pptx):
Slides (pdf):
Recommended Reading:
====================
A Modern Primer on Processing in Memory
Intelligent Architectures for Intelligent Computing Systems
RowHammer: A Retrospective
Fundamentally Understanding and Solving RowHammer
RECOMMENDED LECTURE VIDEOS & PLAYLISTS:
========================================
Computer Architecture Fall 2021 Lectures Playlist:
Digital Design and Computer Architecture Spring 2021 Livestream Lectures Playlist:
Featured Lectures:
Interview with Professor Onur Mutlu:
The Story of RowHammer Lecture:
Accelerating Genome Analysis Lecture:
Memory-Centric Computing Systems Tutorial at IEDM 2021:
Intelligent Architectures for Intelligent Machines Lecture:
Computer Architecture Fall 2020 Lectures Playlist:
Digital Design and Computer Architecture Spring 2020 Lectures Playlist:
Public Lectures by Onur Mutlu, Playlist:
Computer Architecture at Carnegie Mellon Spring 2015 Lectures Playlist:
Rethinking Memory System Design Lecture @stanfordonline :

Comments

Want to join the conversation?

Loading comments...