Julie Grollier
spin torque

bio-inspired hardware need complex nanodevices to be performant

Bio-inspired architectures are composed of processing units interlinked by a huge number of dynamic connections. Nodes and connections perform complex, non-linear operations. In particular, the plastic connections have memory properties, which allow learning and adaptation.

Implementing these functions in CMOS usually requires tens of transistors and external memories, meaning high energy dissipation and large surface on chip, from 10 to 100 Ám2.

This is very problematic, because the network performances increase with the interconnectivity and the number of processing units. For example, in the brain, there are 104 interconnecting synapses per neuron, and more than 1011 neural processing units.

Fabricating non-volatile nanodevices that can emulate the processing units and dynamic connections at the nano-scale therefore appears as the key for the future development of ultra-dense, highly performant bio- inspired chips.