Printed neurons slash AI energy use
- Northwestern University researchers reported on April 15 that they printed flexible artificial neurons from graphene and molybdenum disulfide that fired lifelike spikes. - In Nature Nanotechnology, the devices ran at frequencies up to 20 kilohertz, stayed stable for more than one million cycles, and activated mouse Purkinje neurons. - The work targets neuromorphic chips and brain-machine interfaces, not mainstream AI servers yet. (nature.com)
A neuron is a cell that sends brief electrical pulses, or spikes, instead of doing math the way a graphics processor does. Northwestern University researchers say they have now printed artificial versions of those spiking units on flexible film. (nature.com) (mccormick.northwestern.edu) The study, published April 15 in *Nature Nanotechnology*, came from Mark C. Hersam, Vinod K. Sangwan and colleagues at Northwestern. Their devices use aerosol-jet printing to place graphene and molybdenum disulfide, or MoS2, into fully printed memristors on flexible substrates. (nature.com) (mccormick.northwestern.edu) A memristor is an electrical component whose resistance changes with its history, making it useful for hardware that behaves more like a synapse or neuron than a standard transistor. In this paper, the printed memristors produced threshold switching and spike patterns that the authors describe as first-, second- and third-order spiking complexity. (nature.com) The team reported tunable spiking frequencies up to 20 kilohertz and stable operation over more than 10^6 cycles. They also said the waveforms matched physiological timescales closely enough to stimulate Purkinje neurons in mouse cerebellar slices. (nature.com) That makes this as much a bioelectronics story as a computing story. The paper and Northwestern’s release both frame the devices as candidates for brain-machine interfaces and neuroprosthetics, including systems for hearing, vision and movement. (nature.com) (mccormick.northwestern.edu) The artificial-intelligence angle is more indirect than the headline suggests. The authors argue that brain-like hardware could cut power use by doing some computations with event-driven spikes rather than constant digital processing, but this paper does not report a head-to-head benchmark against commercial graphics processors. (mccormick.northwestern.edu) (nature.com) The paper also positions the work against existing neuromorphic chips that need large numbers of simple artificial neurons for modest tasks. It cites Intel’s Loihi as an example using about 10^6 neurons, while biological circuits can achieve complex behavior with far fewer cells because each neuron is more adaptable. (nature.com) Printed electronics offer a different manufacturing route from conventional silicon: lower-temperature processing, flexible materials and potentially lower cost over large areas. The tradeoff is that lab devices still need scaling, integration and system-level testing before they can compete with established chips. (nature.com) (mccormick.northwestern.edu) So the immediate result is not a drop-in replacement for the hardware training today’s biggest AI models. It is a new printed device platform that can spike like a neuron, survive repeated cycling and send signals into living brain tissue. (nature.com)
