Neuromorphic devices & systems

Technologies for computing tomorrow’s AI

Foundation of new devices & architectures

Many bio-inspired neuromorphic processing units use synthetic neurons, interconnected through variable synaptic weights. To implement the latter, devices with variable “interconnection” strength or plasticity are needed. Moreover, this interconnection strength or “weight” should retain its value for subsequent operations until it is readjusted, for example during the training process. In state-of-the-art production-level neuromorphic processors, both neurons and synaptic weights are implemented as digital circuits or even in software. New devices are needed to map these systems to analog computers with single-device synapses and simple neurons. This requires new materials that are still compatible with legacy CMOS manufacturing processes.

In the Neuromorphic and Devices group at IBM Research – Zurich, we are developing such materials with a focus on transition metal oxides that facilitate device plasticity at the atomic scale through temporal oxygen ion injection, ferroelectric domain manipulation and optically stimulated charge trapping for modulating refractive indices.

 

Intercalation in oxides

Electrical resistive devices arranged as crossbars constitute extremely powerful vector-matrix multipliers and, with additional memory functions, can directly be used as synaptic weights in ANN/CNNs. Thus they help to solve one of the most frequent and most costly operations in typical AI algorithms. These devices require that their programmed resistivity be maintained, which makes them memristive devices. They also require linearity in the transfer and symmetry of their memory and programming characteristics.

In the Neuromorphic and Devices group at IBM Research – Zurich, we are pushing the state-of-the-art of such devices and circuits based on phase-change and ferroelectric effects and with filamentary oxide devices.

Intercalation in oxides

Ferroelectric switching

Ferroelectric domains can be reconfigured in tunnel barriers to alter their electrical conductance. Thus they also act as programmable resistive devices. Our material of choice to exploit this effect is barium titanate (BaTiO3), which we grow epitaxially and hafnium-based compounds such as hafnium zirconium oxide (HZO).

Ferroelectric switching

 

 

 

Ask the experts

Stefan Abel

Stefan Abel
IBM Research scientist

Jean Fompeyrine

Jean Fompeyrine
IBM Research scientist

EU projects

NeuRAM3 logo

NeuRAM3

NEUral computing aRchitectures in Advanced Monolithic 3D-VLSI nano-technologies


plaCMOS logo

plaCMOS

Wafer-scale, CMOS integration of photonics, plasmonics and electronics devices