Tiny Light Circuits Mimic The Brain – At A Fraction Of The Energy Cost

On-chip optical communication between tiny light-based components can make neuromorphic (brain-inspired) computing much smaller and more energy-efficient. In this work, researchers demonstrate that individual nanowire devices on a silicon chip can transmit and receive light signals directly to each other. These miniature circuits communicate reliably, using significantly less power than conventional electronics. The results and models suggest that each operation could use as little as one femtojoule of energy, and that one light source could connect to hundreds of others. This performance meets the requirements for future brain-like networks that can, for example, support autonomous navigation.

In an article published in ACS Photonics, Vidar Flodgren et. al demonstrate optical communication between individual nanowire photodiodes on silicon substrates.

On-chip optical communication between individual nano-optoelectronic components is important to reduce the footprint and improve energy efficiency of photonic neuromorphic solutions. Although nanoscale photon emitters and receivers have been reported separately, communication between them remains largely unexplored. In this work, the researchers demonstrate direct on-chip directional broadcasting of light between individual InP nanowire photodiodes on silicon.

The performance of multiple wire-to-wire communication circuits is mapped, demonstrating robust performance with up to 5 bit resolution as needed in biological networks and a minimum component driving power for continuous operation of 0.5 μW, which is below that of conventional hardware. The results agree well with theoretical modeling that allows the authors to understand network performance limits and identify where significant improvements could be achieved. They estimate that an energy per operation of ∼1 fJ and signal fan-out from one emitter to hundreds of other nodes is possible.

“We find that the nanowire circuit performance parameters can satisfy the quantitative requirements to run the tasks of neural nodes in a bioderived neural network for autonomous navigation”, they write in the abstract.

Co-authors were Abhijit Das, Joachim E. Sestoft, David Alcer, Thomas K. Jensen, Hossein Jeddi, Håkan Pettersson, Jesper Nygård, Magnus T. Borgström, Heiner Linke, and Anders Mikkelsen.