ICE seminar: « Memory-centric computing »

DALGATY_telecom_compressedAbstract

Computing transforms data into useful information. To compute on data, it must be transported from a sensor where it is generated, or from a memory where it is stored, to a circuit where it is operated on. The rapid and constant charging and discharging of metal wires to transmit this information is the major latency, bandwidth and energetic bottleneck in today’s computing systems. To design faster and more efficient computers, one solution is to bring computation near — or even inside — the sensor and the memory. However, constraints related to form factor, device physics, data representation, and algorithm design make this extremely challenging. This talk will give an overview of the three principal means explored at the CEA to build more memory-centric computers by re-thinking – (i) algorithms in order to reduce redundancy in memory, (ii) computer architectures, to minimize the physical distance between memory and operators and (iii) computing with the analogue characteristics of memory devices so that memory itself becomes the operator.

Bio

Thomas obtained a Master’s degree in Electrical Engineering from the University of Glasgow in 2017, followed by a PhD conducted between CEA-Leti in Grenoble and the Institute of Neuroinformatics in Zurich. Since 2020, he has been a researcher at CEA-List in Grenoble, and in 2024 he was recognized as a CEA expert in emerging computing paradigms. His work has been published in Nature Electronics/Materials/Communications as well as NeurIPS,and CVPR and he has deposited thirty patents.