Training today’s largest AI models requires enormous data centres filled with processors that consume power around the clock. Some analyses predict that, in the coming years, AI could account for a significant share of the world’s growing electricity consumption.
Meanwhile, the human brain performs tasks that still challenge even the most advanced AI systems while consuming just 20 watts of power – roughly the same as an LED bulb.
“The transistors we use today are actually incredibly energy-efficient. The problem is that information has to be moved back and forth between them and the memory all the time. It is the transport of information that consumes energy. The brain works differently. Here, computation, memory and communication take place within the same networks of connections,” says Anders Mikkelsen, who led the new study together with colleagues from Lund University in Sweden and the University of Copenhagen in Denmark.
This is why researchers around the world are looking for new ways to build AI. If AI models continue to grow at the same pace as they do today, energy consumption could become one of the biggest constraints on how far the technology can develop.
In a new study, researchers from Lund University and the University of Copenhagen have developed an artificial neuron that brings together several of the functions researchers have long been seeking in photonic neurons.
- Read the entire article by Morten Busch on the Science News website
- “Nanoscale photonic artificial neuron with biological signal processing” in Nature Communications
- Anders Mikkelsen’s profile in the Lund University Research Portal
Authors: Joachim E. Sestoft, Thomas K. Jensen, Vidar Flodgren, Abhijit Das, Rasmus D. Schlosser, David Alcer, Mariia Lamers, Thomas Kanne, Magnus T. Borgström, Jesper Nygård & Anders Mikkelsen
From the abstract:
“The rising energy demand of artificial intelligence infrastructure is not sustainable. Neuromorphic hardware offers encouraging solutions to this problem by mimicking the energy-efficient biological brain. Many different hardware solutions have been suggested, but photonic components are especially promising in terms of speed and power-efficiency and similar computing hardware can also serve as optical signaling/sensory systems. However, many still lack one or more essentials, like: (i) miniaturized building blocks for high density integration; (ii) excitation and inhibition in the same device; (iii) linear fan-in (summation) and tunable nonlinear activation; (iv) low optical energy per operation; (v) controlled device-to-device variation and simple tunable weighting; (vi) wavelength selectivity for routing; and (vii) CMOS-compatible materials and processing, with a clear path to all-optical and on-chip links. For optical sensory systems many of the same demands are highly relevant in order to mimic the exceptional analytical power of the biological retina. This includes contrast resolution over many orders of magnitude of background light, excellent dynamic range and edge resolution. For intensity adaptation and edge sharpening inhibition plays an important role.
Here we combine three semiconductor nanowires to construct an artificial optical/electronic (O/E) neuron, that fulfills these requirements. The active area is 30-90 μm2 (at least 100 times smaller than prior on-chip photonic activators) it provides both excitation and inhibition, sums concurrent optical inputs, and provides sigmoid activation functionality. It operates at pico-watt optical powers, shows millisecond-scale responses with 0.1-1 s recovery, is responsive across multiple wavelengths, offers voltage-tunable sensitivity/weighting and modelling projects that it can support ~ 1 GHz operation speeds. Compared with planar photonic platforms, our nanowire node provides an exceptionally high absorption cross-section per footprint together with bandgap-, geometry- and orientation-controlled wavelength and polarization selectivity. Nanowire technologies are highly refined and employed in many different technology areas (e.g. quantum computing and solar cells) and the nodal architecture allows for combinations of diverse functionalities in a modular fashion, while being CMOS-compatible.”