Building energy-efficient and reliable systems-on-chip for AI processing: our recent experience with mixed-signal compute-in-memory AI accelerators

Jeudi 27 mars 2025
Prof. Martin Andraud, UCLouvain, Belgium

The necessity for building always more energy-efficient processing systems dedicated to AI tasks Towards appears more evident every day, pushed by the ever-increasing need for computation of AI models. Towards this goal, Compute In Memory (CIM) is becoming an architecture of reference to accelerate AI computation. Although we can build CIM computing core digitally, achieving further 10-100x improvements in energy efficiency, necessary for the next generation of embedded AI systems, may rely on analog computing principles. In particular, emerging Non-Volatile Memories (eNVMs) offer an alternative way to store and compute AI data more efficiently, breaking the traditional cost-efficiency-volatility trade-offs in memory design. Yet, building complete AI systems integrating reliable mixed-signal CIM computing cores is not straightforward.

Examples of these challenges are:

integrating CIM cores into complete systems, for instance, with a digital control processor,
limiting the variability of analog circuitries to ensure sufficient computation accuracy, or
allowing to automatically program and test the system.

Hence, in this talk, I will walk you through how my group approaches the design of analog CIM circuits, focusing on these practical aspects that are typically not fully explained. Examples of challenges we face(d) are designing an effective analog computing cell, designing circuit peripherals, integrating the CIM core into a full system, and ensuring more reliability and accuracy in the system through self-calibration.

Speaker's Biography: Martin Andraud is an assistant professor in microelectronics at UCLouvain, Belgium, and a visiting professor at Aalto University, Finland. He received his PhD from Grenoble University, France, in 2016. He was a postdoctoral researcher successively with TU Eindhoven in 2016 and KU Leuven from 2017 to 2019. Then, he was an assistant professor at Aalto University, Finland, from 2029 to 2023. His research interests include ASIC design for hybrid AI tasks (e.g., deep learning, neurosymbolic AI, or probabilistic AI), hardware/software co-design of AI accelerator systems, as well as test and reliability of custom ASICs for digital or mixed-signal AI acceleration.