Experimenting with neural architectures in silicon
The NeuRAM3 project (neural-computing architectures in advanced monolithic 3D VLSI technologies) is a three-year EU research program including teams from CEA Tech institutes Leti and List (F), STMicroelectronics (F), IBM Zurich (SUI), University of Zurich (SUI), CNR-IMM (I), imec (NL, B), Jacobs University (D), and IMSE-CISC (ESP).
A year in the making, the project participants will present their latest finding both on the materials side and on the architectural side, exploring various design routes to implement neuromorphic circuit blocks with machine-learning algorithms and spike-based learning mechanisms that could one day complement or even displace traditional computer architectures.
With integration in mind, the project’s goals include the development of an ultra-low power, scalable and highly reconfigurable neural architecture, with a 50x improvement in power consumption compared to conventional digital solutions. The architecture should lend itself to 3D monolithic integration using ST’s FDSOI process at 28nm, with integrated RRAM synaptic elements.
“We are on track with our planning, on the materials side, we are finalizing our choices, with test results on OxRAM and CBRAM materials, and we are busy modelling these synaptic memories”, told us Carlo Reita, director technical marketing and strategy, nanoelectronics at Leti.
“In the next few months, we’ll feedback our results into design activities, concentrating on a chip design on a Multi-Project Wafer run” Reita continued.
“With our first tape out, we hope to validate a neuromorphic architecture, and in a second phase, we would want to replace all SRAM on the chip with resistive RAM”.
Reading from some of the papers to be presented, neuromorphic machines will soon make CAPTCHA tricks obsolete.
Register for the workshop at https://esscirc-essderc2016.epfl.ch/registration
Visit Leti at www.leti.fr