All-optical computing with beyond 100-GHz clock rates

Abstract: A computer's clock rate ultimately determines the minimum time between sequential operations or instructions. Despite exponential advances in electronic computer performance owing to Moore's Law and increasingly parallel system architectures, computer clock rates have remained stagnant at $\sim5~\mathrm{GHz}$ for almost two decades. This poses an intractable problem for applications requiring real-time processing or control of ultrafast information systems. Here we break this barrier by proposing and experimentally demonstrating computing based on an end-to-end and all-optical recurrent neural network harnessing the ultrafast nature of linear and nonlinear optical operations while avoiding electronic operations. The all-optical computer realizes linear operations, nonlinear functions, and memory entirely in the optical domain with $>100~\mathrm{GHz}$ clock rates. We experimentally demonstrate a prototypical task of noisy waveform classification as well as perform ultrafast in-situ analysis of the soliton states from integrated optical microresonators. We further illustrate the application of the architecture for generative artificial intelligence based on quantum fluctuations to generate images even in the absence of input optical signals. Our results highlight the potential of all-optical computing beyond what can be achieved with digital electronics by utilizing ultrafast linear, nonlinear, and memory functions and quantum fluctuations.

• The paper presents an AO-RNN that achieves clock speeds over 100 GHz, surpassing the limitations of electronic processors.
• The methodology leverages time-multiplexed photonic setups with Mach-Zehnder interferometers and nonlinear waveguides for full optical processing.
• The experimental results show effective noisy waveform classification and regression performance, underscoring its potential for real-time optical computing.

All-optical Computing with Beyond 100-GHz Clock Rates

This paper presents a notable advancement in optical computing, experimentally demonstrating an all-optical recurrent neural network (AO-RNN) achieving clock rates exceeding 100 GHz. This represents a significant improvement over the stagnant growth of electronic computer clock rates, which have plateaued around 5 GHz for nearly two decades due to limitations in transistor scaling and architectural bottlenecks. The introduction of the AO-RNN leverages the ultrafast properties of optical components, namely linear and nonlinear optical operations, to surpass the speed constraints of electronic systems.

Experimental Framework

The proposed AO-RNN operates entirely within the optical domain, performing linear, nonlinear, and memory operations without relying on electronic modulation or computation. The architecture is based on a recurrent neural network model, suitable for handling temporal or sequential data. The AO-RNN utilizes time-multiplexed photonic configurations and high-repetition-rate ultrashort laser pulses to achieve rapid analogue processing, a notable deviation from conventional digital processors constrained by electronic clock rates.

The network architecture encompasses interconnecting components like multi-arm Mach-Zehnder interferometers for linear operations, periodically poled lithium niobate waveguides for ultrafast nonlinear activations, and active optical cavities for memory functions. This design enables the network to process data and generate outputs at a rate dictated by the laser pulse repetition, effectively setting the clock rate of the optical computer.

Key Findings and Implications

1. Noisy Waveform Classification: The AO-RNN was applied to classify noisy waveforms at clock rates beyond 100 GHz. It managed a classification accuracy of 58% even at maximum tested rates (100 GHz), surpassing random guess accuracy and showcasing the functional role of ultrafast nonlinear optics in this regime.
2. Native Ultrafast Optical Handling: Utilizing the AO-RNN for analyzing ultrafast signals from optical microresonators demonstrated its capability to function as an in-situ tool for real-time ultrafast optical signal characterization, a feat unattainable by electronic systems.
3. Regression Tasks: In forecasting future values in time-series data, the AO-RNN achieved low normalized mean square errors, highlighting its potential for high-speed continuous data processing, an area pivotal for real-time decision-making applications across diverse domains.
4. Generative Applications: The integration of quantum fluctuations as seeds in a generative model presents an innovative use of AO-RNNs for application in generating handwritten digit images without input signals.

Future Prospects and Developments

The paper suggests substantial room for enhancing the AO-RNN's performance through integration with advanced photonic technologies, such as thin-film lithium niobate, to realize faster modulators and switches. The transition to integrated photonic platforms could also significantly reduce latency and physical size, paving the way for more widespread utilization of optical neural networks in practical applications.

From a theoretical standpoint, these developments bolster optical computing's role in mordern computation paradigms, particularly in scenarios demanding real-time processing capabilities. The demonstrated AO-RNN highlights the potential for optical networks to complement or substitute electronic systems in high-speed digital and analog computing tasks, indicating a promising trajectory toward faster, more efficient computational architectures.