23 September 2022 in Nature Communications
The explosive growth of deep learning applications has triggered a new era in computing hardware, targeting the efficient deployment of multiply-and-accumulate operations. In this realm, integrated photonics have come to the foreground as a promising energy efficient deep learning technology platform for enabling ultra-high compute rates. However, despite integrated photonic neural network layouts have already penetrated successfully the deep learning era, their compute rate and noise-related characteristics are still far beyond their promise for high-speed photonic engines. Herein, we demonstrate experimentally a noise-resilient deep learning coherent photonic neural network layout that operates at 10GMAC/sec/axon compute rates and follows a noise-resilient training model. The coherent photonic neural network has been fabricated as a silicon photonic chip and its MNIST classification performance was experimentally evaluated to support accuracy values of >99% and >98% at 5 and 10GMAC/sec/axon, respectively, offering 6× higher on-chip compute rates and >7% accuracy improvement over state-of-the-art coherent implementations.
21 February 2020 in IEEE Journal of Selected Topics in Quantum Electronics
Photonic artificial neural networks have garnered enormous attention due to their potential to perform multiply-accumulate (MAC) operations at much higher clock rates and consuming significantly lower power and chip real-estate compared to digital electronic alternatives. Herein, we present a comprehensive power consumption analysis of photonic neurons, taking into account global design parameters and concluding to analytical expressions for the neuron's energy- and footprint efficiencies. We identify the optimal design-space and analyze the performance plateaus and their dependence on a range of physical parameters, highlighting the existence of an optimal data-rate for maximizing the energy efficiency. Following a survey of the best-in-class integrated photonic devices, including on-chip lasers, photodetectors, modulators and weighting elements, the mathematically calculated energy and footprint efficiencies are mapped into real photonic neuron deployment scenarios. We reveal that silicon photonics can compete with the best-performing currently available digital electronic neural network engines, reaching ${\rm{TMAC/s/mm^{2}}}$ footprint- and sub-pJ/MAC energy efficiencies. Simultaneously, neuromorphic plasmonics, plasmo-photonics and sub-wavelength photonics hold the credentials for 1 to 3 orders of magnitude improvements even when the laser requirements and a reasonable waveguide pitch are accounted for, promising performance at a few fJ/MAC and up to a few ${\rm{TMAC/s/mm^{2}}}$ .
3 May 2022 in Journal of Lightwave Technology
The relentless growth of Artificial Intelligence (AI) workloads has fueled the drive towards non-Von Neuman architectures and custom computing hardware. Neuromorphic photonic engines aspire to synergize the low-power and high-bandwidth credentials of light-based deployments with novel architectures, towards surpassing the computing performance of their electronic counterparts. In this paper, we review recent progress in integrated photonic neuromorphic architectures and analyze the architectural and photonic hardware-based factors that limit their performance. Subsequently, we present our approach towards transforming silicon coherent neuromorphic layouts into high-speed and high-accuracy Deep Learning (DL) engines by combining robust architectures with hardware-aware DL training. Circuit robustness is ensured through a crossbar layout that circumvents insertion loss and fidelity constraints of state-of-the-art linear optical designs. Concurrently, we employ DL training models adapted to the underlying photonic hardware, incorporating noise- and bandwidth-limitations together with the supported activation function directly into Neural Network (NN) training. We validate experimentally the high-speed and high-accuracy advantages of hardware-aware DL models when combined with robust architectures through a SiPho prototype implementing a single column of a 4:4 photonic crossbar. This was utilized as the pen-ultimate hidden layer of a NN, revealing up to 5.93% accuracy improvement at 5GMAC/sec/axon when noise-aware training is enforced and allowing accuracies of 99.15% and 79.8% for the MNIST and CIFAR-10 classification tasks. Channel-aware training was then demonstrated by integrating the frequency response of the photonic hardware in NN training, with its experimental validation with the MNIST dataset revealing an accuracy increase of 12.93% at a record-high rate of 25GMAC/sec/axon.
1 June 2022 in Optical Materials Express
We review different technologies and architectures for neuromorphic photonic accelerators, spanning from bulk optics to photonic-integrated-circuits (PICs), and assess compute efficiency in OPs/Watt through the lens of a comparative study where key technology aspects are analyzed. With an emphasis on PIC neuromorphic accelerators, we shed light onto the latest advances in photonic and plasmonic modulation technologies for the realization of weighting elements in training and inference applications, and present a recently introduced scalable coherent crossbar layout. Finally, we stress that current technologies face challenges endowing photonic accelerators with compute efficiencies in the PetaOPs/W, and discuss future implementation pathways towards improving performance.
6 September 2018 in Scientific Reports
ECEvangelia Chatzianagnostou
Evangelia Chatzianagnostou
+8
Co-integrating CMOS plasmonics and photonics became the “sweet spot” to hit in order to combine their benefits and allow for volume manufacturing of plasmo-photonic integrated circuits. Plasmonics can naturally interface photonics with electronics while offering strong mode confinement, enabling in this way on-chip data interconnects when tailored to single-mode waveguides, as well as high-sensitivity biosensors when exposing Surface-Plasmon-Polariton (SPP) modes in aqueous environment. Their synergy with low-loss photonics can tolerate the high plasmonic propagation losses in interconnect applications, offering at the same time a powerful portfolio of passive photonic functions towards avoiding the use of bulk optics for SPP excitation and facilitating compact biosensor setups. The co-integration roadmap has to proceed, however, over the utilization of fully CMOS compatible material platforms and manufacturing processes in order to allow for a practical deployment route. Herein, we demonstrate for the first time Aluminum plasmonic waveguides co-integrated with Si3N4 photonics using CMOS manufacturing processes. We validate the data carrying credentials of CMOS plasmonics with 25 Gb/s data traffic and we confirm successful plasmonic propagation in both air and water-cladded waveguide configurations. This platform can potentially fuel the deployment of co-integrated plasmonic and photonic structures using CMOS processes for biosensing and on-chip interconnect applications.
19 May 2020 in IEEE Journal of Selected Topics in Quantum Electronics
Neuromorphic photonics came to the fore promising neural networks (NNs) with orders of magnitude higher computational speeds compared to electronic counterparts. In this direction, research efforts have been mainly concentrated on the development of spiking, convolutional and Feed-Forward (FF)-NN architectures, aiming to solve complex cognitive problems. However, in order to solve time-series classification and prediction complex tasks, state-of-the-art deep-learning models require in most cases the employment of Recurrent-NNs (RNNs) along with their gated variants, such as Long-Short-Term-Memories (LSTMs) and Gated-Recurrent-Units (GRUs). Herein, we experimentally demonstrate the first, to the best of our knowledge, all-optical RNN with a gating mechanism, laying the foundations for all-optical LSTMs and GRUs. The proposed layouts exploit a Semiconductor-Optical-Amplifier (SOA)-based sigmoid activation within a fiber loop and were validated using asynchronous Wavelength--(WDM) signals with 100psec optical pulses. A SOAMach-Zehnder-Interferometer (SOA-MZI) gate was employed in the Gated-RNN version, with the RNN output defining the input signal fraction that is desired to enter the RNN. Finally, a complex NN architecture was trained using the FI-2010 financial dataset exploiting the proposed non-gated and gated-RNNs, showcasing in an outstanding F1 score of 41.68% and 41.85%, respectively, outperforming the Multi-Layer Perceptron (MLP) based models by 6.49% in average.