In-memory computing to solve AI’s energy consumption bottle-neck
Recording from AI Lund lunch seminar 5 May 2021
Ferroelectric memristors for ultra-low power neuromorphic computing.
Title: In-memory computing to solve AI’s energy consumption bottle-neck
When: 5 May at 12.00-13.15
Speaker: Mattias Borg, Dept. of Electrical and Information Technology, LTH , Lund University
The bottle-neck for continued development of Machine Learning lies in the escalating energy consumption during model training. Ultimately, this will require new hardware that implements non-von Neumann architectures, enabling computing-in-memory and even online unsupervised learning by brain-inspired methods.
Memristors based on ferroelectric memory elements are a promising route to such hardware. Here I will introduce memristor-based computing-in-memory, it's benefits in terms of energy-efficiency and our research on the ferroelectric devices that can make it reality.