image: Figure 1 | Vision of the project NEHO: Neuromorphic computing Enabled by Heavily doped semiconductor Optics. On the left panel a ring resonator evanescently couples to two ridge waveguides (blue layers): the two propagating modes in the ridge waveguides interact through the ring resonator, where the nonlinearity makes the degree of interaction dependent on the input intensity of the propagating modes (signal strength). This typical photonic integrated chip structure can be adapted to the NEHO concept of a reconfigurable optical nonlinearity based on free carrier hydrodynamics in a thin semiconductor layer that we demonstrate in this work. In the future, the free carrier density could be further tuned by a field effect gate (yellow layer) making the nonlinear coefficient reconfigurable at will, as required by modern neuromorphic computing approaches (right panel).
Credit: Andrea Rossetti, Huatian Hu et al.
Deep learning applications are expected to add trillions of dollars annually to the global economy. Expenses associated with state-of-the-art artificial intelligence (AI) model training have been increasing at a rate of approximately 2.4 times per year since 2016. At this pace, the cost to train the most compute-intensive AI models could exceed $1 billion by 2027. This cost growth will equally increase the power consumption needs and their impact on global CO2 production. For example, training OpenAI's GPT-3 model consumed approximately 1,287 megawatt-hours (MWh) of electricity, which is about as much as the average U.S. household uses over 120 years.
It is then beyond any doubt that there is a concrete need for alternative schemes to do AI-specialized high-performance low-power-consumption signal processing.
In a new paper published in Light: Science & Applications, a team of scientists, led by Dr. Cristian Ciracì of Istituto Italiano di Tecnologia, now at the company Neurophos in the US, and Prof. Michele Ortolani of Sapienza University of Rome, Italy demonstrated the fundamental principles of a new technology that embeds the efficient electromagnetic nonlinearities associated with free-carrier dynamics in heavily doped semiconductors directly into chips that can be produced with integrated optical circuit design.
The work is part of the EU-funded project NEHO, now led by Valeria Giliberti of Istituto Italiano di Tecnologia (IIT) in Italy, which aims to develop a revolutionary technology allowing all-optical low-power-consumption AI training for next generation data centers. By leveraging the unconventional use of semiconductors to optimize and control plasmonic effects, NEHO seeks to implement an integrated optical neuron as the building block of ultrafast neural networks (NNs).
The scientists have experimentally demonstrated that ultrafast optical nonlinearities in doped semiconductors can be engineered and can easily exceed those of conventional undoped dielectrics. The electron response of heavily doped semiconductors acquires in fact a hydrodynamic character that introduces nonlocal effects as well as additional nonlinear sources. The team’s experimental findings are supported by a comprehensive computational analysis based on the hydrodynamic model. In particular, by studying third-harmonic generation from plasmonic nanoantenna arrays made out of heavily n-doped InGaAs with increasing levels of free-carrier density, it is possible to discriminate between hydrodynamic and dielectric nonlinearities. Most importantly, the authors demonstrate that the maximum nonlinear efficiency as well as its spectral location can be engineered by tuning the doping level. Crucially, the maximum efficiency can be increased by almost two orders of magnitude with respect to the classical dielectric nonlinearity. Having employed the common material platform InGaAs/InP that supports integrated waveguides, these findings pave the way for future exploitation of plasmonic nonlinearities in all-semiconductor photonic integrated circuits.
Journal
Light Science & Applications
Article Title
Control and enhancement of optical nonlinearities in plasmonic semiconductor nanostructures