Edge AI with In-memory Computing
Team
Sources of Funding
| Cerberus |  |
The introduction of the on-device AI paradigm in edge computing could bring significant benefits and enable a plethora of exciting applications in AR/VR gear, industrial automation devices, drones, 5G/6G base stations or autonomous vehicles. However, AI is well- known to be generally power hungry and edge devices are not able to withstand the computational requirements of emerging AI workloads. Therefore, energy efficiency is closing the door important breakthroughs across industries such as ICT, mobility, healthcare, agri-food, industrial automation, leaving an enormous untapped potential behind. To remedy this, academia and industry have proposed in-memory computing (IMC) and the associated non-Von Neumann architectures as a means to dramatically improve the efficiency of AI processing; yet the promise of the IMC paradigm is limited by non-idealities within and around the IMC tiles.
In this light, the CERBERUS project aims to take a radical step forward in the field of IMC by addressing the fundamental efficiency bottlenecks that prevent edge computing stakeholders from unleashing the full potential of this technology. To this end, CERBERUS will employ a three-pillar approach grounded on:
(i) 2× more efficient memristive devices,
(ii) 30% more efficient peripheral circuits also allowing to bypass domain conversions, and
(iii) 70% more effective heat dissipation devices to virtually eliminate thermal noise.
Thanks to a deeply cross-layer simulation approach integrating models of the developed devices and circuits, our ultimate goal is to show that 3X better energy efficiency can be achieved at the system level for modern and emerging edge AI computing workloads, towards the vision of deploying sizable AI models in virtually any device, even at the extreme edge. |
News
Related Publications
