TDK turned to the CEA-List to help expand into AI and, in particular, edge learning, with its magnetic spin memristor technology. The challenge is memristors’ intrinsic noise, which makes them difficult to program accurately. Even with costly variability reduction strategies, standard machine learning techniques produce sub-optimal results.
Instead of trying to eliminate noise, the CEA-List’s research engineers took a different approach: They decided to use the noise inherent to memristors to their advantage. Since 2020, the CEA-List has been pioneering Bayesian in-memory computing, an approach that uses memristors’ randomness to implement efficient probabilistic algorithms. We developed a programming strategy that leverages an advanced probabilistic method and the physics of TDK’s memory devices, achieving power and latency improvements of two to three orders of magnitude over today’s digital hardware. Model sizes could be increased fivefold with exceptionally low power consumption on this new type of hardware.
This disruptive approach turned a limitation into a powerful strength, revolutionizing how we think about computing for the AI of the future. The partnership between the CEA and TDK has earned international recognition in the form of a best paper award at the NeurIPS conference. Two patents have also been filed.
power efficiency gains, a disruptive approach to AI computation models
Contributed to this article:
Bayesian in-memory computing is an emerging field. This research will create new avenues to more sustainable, reliable, and efficient solutions to meet the growing demands of modern artificial intelligence applications.
This new development by the CEA opens up new applications for the spin memristor, and represents an advance toward future ultra-low-power AI chips—a goal shared by both organizations.