The wide adoption and significant computing resource cost of attention-based transformers, e.g., Vision Transformers and large language models, have driven the demand for efficient hardware accelerators. While electronic accelerators have been commonly used, there is a growing interest in exploring photonics as an alternative technology due to its high energy efficiency and ultra-fast processing speed. Photonic accelerators have demonstrated promising results for convolutional neural networks (CNNs) workloads, which predominantly rely on weight-static linear operations. However, they encounter challenges when it comes to efficiently supporting attention-based Transformer architectures, raising questions about the applicability of photonics to advanced machine-learning tasks. The primary hurdle lies in their inefficiency in handling the unique workloads inherent to Transformers, i.e., dynamic and full-range tensor multiplication. In this work, we propose Lightening-Transformer, the first light-empowered, high-performance, and energy-efficient photonic Transformer accelerator. To overcome the fundamental limitation of existing photonic tensor core designs, we introduce a novel dynamically-operated photonic tensor core, DPTC, consisting of a crossbar array of interference-based optical vector dot-product engines, supporting highly parallel, dynamic, and full-range matrix multiplication. Furthermore, we design a dedicated accelerator that integrates our novel photonic computing cores with photonic interconnects for inter-core data broadcast, fully unleashing the power of optics. The comprehensive evaluation demonstrates that Lightening-Transformer achieves >2.6x energy and > 12 x latency reductions compared to prior photonic accelerators and delivers the lowest energy cost and 2 to 3 orders of magnitude lower energy-delay product compared to the electronic Transformer accelerator, all while maintaining digital-comparable accuracy. Our work highlights the immense potential of photonics for efficient hardware accelerators, particularly for advanced machine-learning workloads, such as Transformer-backboned large language models (LLM). Our implementation is available at https://github.com/zhuhanqing/Lightening-Transformer.

Optical neural networks (ONNs) are promising hardware platforms for next-generation artificial intelligence acceleration with ultrafast speed and low-energy consumption. However, previous ONN designs are bounded by one multiply–accumulate operation per device, showing unsatisfying scalability. In this work, we propose a scalable ONN architecture, dubbedSqueezeLight. We propose a nonlinear optical neuron based on multioperand ring resonators (MORRs) to squeeze vector dot-product into a single device with low wavelength usage and built-in nonlinearity. A block-level squeezing technique with structured sparsity is exploited to support higher scalability. We adopt a robustness-aware training algorithm to guarantee variation tolerance. To enable a truly scalable ONN architecture, we extendSqueezeLightto a separable optical CNN architecture that further squeezes in the layer level. Two orthogonal convolutional layers are mapped to one MORR array, leading to order-of-magnitude higher software training scalability. We further explore augmented representability forSqueezeLightby introducing parametric MORR neurons with trainable nonlinearity, together with a nonlinearity-aware initialization method to stabilize convergence. Experimental results show thatSqueezeLightachieves one-order-of-magnitude better compactness and efficiency than previous designs with high fidelity, trainability, and robustness. Our open-source codes are available athttps://github.com/JeremieMelo/SqueezeLight.

Optical phase change material (PCM) has emerged promising to enable photonic in-memory neurocomputing in optical neural network (ONN) designs. However, massive photonic tensor core (PTC) reuse is required to implement large matrix multiplication due to the limited single-core scale. The resultant large number of PCM writes during inference incurs serious dynamic energy costs and overwhelms the fragile PCM with limited write endurance, causing the severe aging issue. Moreover, the aged PCM would distort the stored value and significantly degrade the reliability of PTC. In this work, we propose a holistic solution,ELight, to tackle both the aging issue and the post-aging reliability issue, where a proactive aging-aware optimization framework minimizes the overall PCM write cost and a post-aging tolerance scheme overcomes the effect of aged PCM. Specifically, in the aging-aware optimization part, we propose write-aware training to encourage the similarity among weight blocks and combine it with a post-training optimization technique to reduce programming efforts by eliminating redundant writes. Next, an efficient groupwise row-based weight-PTC remapping scheme is introduced to tolerate the reprogrammability degradation due to the aged PCM. Experiments show thatELightcan achieve over$20 \times $reductions in the total number of write operations and dynamic energy cost with comparable accuracy. Moreover,ELightcan guarantee significant accuracy recovery under the aged PCM within photonic memories. With ourELight, photonic in-memory neurocomputing will step forward toward practical applications in machine learning with order-of-magnitude longer lifetime, lower programming energy cost, and significant resilience against PCM aging effects.

With the recent advances in optical phase change material (PCM), photonic in-memory neurocomputing has demonstrated its superiority in optical neural network (ONN) designs with near-zero static power consumption, time-of-light latency, and compact footprint. However, photonic tensor cores require massive hardware reuse to implement large matrix multiplication due to the limited single-core scale. The resultant large number of PCM writes leads to serious dynamic power and overwhelms the fragile PCM with limited write endurance. In this work, we propose a synergistic optimization framework, ELight, to minimize the overall write efforts for efficient and reliable optical in-memory neurocomputing. We first propose write-aware training to encourage the similarity among weight blocks, and combine it with a post-training optimization method to reduce programming efforts by eliminating redundant writes. Experiments show that ELight can achieve over$20\times$reduction in the total number of writes and dynamic power with comparable accuracy. With our ELight, photonic in-memory neurocomputing will step forward towards viable applications in machine learning with preserved accuracy, order-of-magnitude longer lifetime, and lower programming energy.

Photonic tensor cores (PTCs) are essential building blocks for optical artificial intelligence (AI) accelerators based on programmable photonic integrated circuits. PTCs can achieve ultra-fast and efficient tensor operations for neural network (NN) acceleration. Current PTC designs are either manually constructed or based on matrix decomposition theory, which lacks the adaptability to meet various hardware constraints and device specifications. To our best knowledge, automatic PTC design methodology is still unexplored. It will be promising to move beyond the manual design paradigm and "nurture" photonic neurocomputing with AI and design automation. Therefore, in this work, for the first time, we propose a fully differentiable framework, dubbed ADEPT, that can efficiently search PTC designs adaptive to various circuit footprint constraints and foundry PDKs. Extensive experiments show superior flexibility and effectiveness of the proposed ADEPT framework to explore a large PTC design space. On various NN models and benchmarks, our searched PTC topology outperforms prior manually-designed structures with competitive matrix representability, 2×-30× higher footprint compactness, and better noise robustness, demonstrating a new paradigm in photonic neural chip design. The code of ADEPT is available at link using the TorchONN library.

Optical computing has become emerging technology in next-generation efficient artificial intelligence (AI) due to its ultra-high speed and efficiency. Electromagnetic field simulation is critical to the design, optimization, and validation of photonic devices and circuits.However, costly numerical simulation significantly hinders the scalability and turn-around time in the photonic circuit design loop. Recently, physics-informed neural networks were proposed to predict the optical field solution of a single instance of a partial differential equation (PDE) with predefined parameters. Their complicated PDE formulation and lack of efficient parametrization mechanism limit their flexibility and generalization in practical simulation scenarios. In this work, for the first time, a physics-agnostic neural operator-based framework, dubbed NeurOLight, is proposed to learn a family of frequency-domain Maxwell PDEs for ultra-fast parametric photonic device simulation. Specifically, we discretize different devices into a unified domain, represent parametric PDEs with a compact wave prior, and encode the incident light via masked source modeling. We design our model to have parameter-efficient cross-shaped NeurOLight blocks and adopt superposition-based augmentation for data-efficient learning. With those synergistic approaches, NeurOLight demonstrates 2-orders-of-magnitude faster simulation speed than numerical solvers and outperforms prior NN-based models by ~54% lower prediction error using ~44% fewer parameters.

Reconfigurable intelligent surface (RIS)-assisted millimeter-wave (mmWave) communication systems relying on hybrid beamforming structures are capable of achieving high spectral efficiency at a low hardware complexity and low power consumption. In this article, we propose an RIS-assisted mmWave point-to-point system relying on dynamically configured subarray connected hybrid beamforming structures. More explicitly, an energy-efficient analog beamformer relying on the twin-resolution phase shifters is proposed. Then, we conceive a successive interference cancelation (SIC)-based method for jointly designing the hybrid beamforming matrix of the base station (BS) and the passive beamforming matrix of the RIS. Specifically, the associated bandwidth-efficiency maximization problem is transformed into a series of subproblems, where the subarray of phase shifters and RIS elements is jointly optimized for maximizing each subarray’s rate. Furthermore, a greedy method is proposed for determining the phase shifter configuration of each subarray. We then propose to update the RIS elements relying on a complex circle manifold (CCM)-based method. The proposed dynamic subconnected structure as well as the proposed joint hybrid and passive beamforming method strike an attractive tradeoff between the bandwidth efficiency and power consumption. Our simulation results demonstrate the superiority of the proposed method compared to its traditional counterparts.

The cell-free MIMO concept relying on hybrid precoding constitutes an innovative technique capable of dramatically increasing the network capacity of millimeter-wave (mmWave) communication systems. It dispenses with the cell boundary of conventional multi-cell MIMO systems, while drastically reducing the power consumption by limiting the number of radio frequency (RF) chains at the access points (APs). In this paper, we aim for maximizing the weighted sum rate (WSR) of mmWave cell-free MIMO systems by conceiving a low-complexity hybrid precoding algorithm. We formulate the WSR optimization problem subject to the transmit power constraint for each AP and the constant-modulus constraint for the phase shifters of the analog precoders. A block coordinate descent (BCD) algorithm is proposed for iteratively solving the problem. In each iteration, the classic Lagrangian multiplier method and the penalty dual decomposition (PDD) method are combined for obtaining near-optimal hybrid analog/digital precoding matrices. Furthermore, we extend our proposed algorithm for deriving closed-form expressions for the precoders of fully digital cell-free MIMO systems. Moreover, we present the convergency analysis and complexity analysis of our proposed method. Finally, our simulation results demonstrate the superiority of the algorithms proposed for both fully digital and hybrid precoding matrices.

Optical neural networks (ONNs) have demonstrated record-breaking potential in high-performance neuromorphic computing due to their ultra-high execution speed and low energy consumption. However, current learning protocols fail to provide scalable and efficient solutions to photonic circuit optimization in practical applications. In this work, we propose a novel on-chip learning framework to release the full potential of ONNs for power-efficient in situ training. Instead of deploying implementation-costly back-propagation, we directly optimize the device configurations with computation budgets and power constraints. We are the first to model the ONN on-chip learning as a resource-constrained stochastic noisy zeroth-order optimization problem, and propose a novel mixed-training strategy with two-level sparsity and power-aware dynamic pruning to offer a scalable on-chip training solution in practical ONN deployment. Compared with previous methods, we are the first to optimize over 2,500 optical components on chip. We can achieve much better optimization stability, 3.7x-7.6x higher efficiency, and save >90% power under practical device variations and thermal crosstalk.

Optical neural networks (ONNs) have demonstrated promising potentials for next-generation artificial intelligence acceleration with ultra-low latency, high bandwidth, and low energy consumption. However, due to high area cost and lack of efficient sparsity exploitation, previous ONN designs fail to provide scalable and efficient neuromorphic computing, which hinders the practical implementation of photonic neural accelerators. In this work, we propose a novel design methodology to enable a more scalable ONN architecture. We propose a nonlinear optical neuron based on multi-operand ring resonators to achieve neuromorphic computing with a compact footprint, low wavelength usage, learnable neuron balancing, and built-in nonlinearity. The structured sparsity is exploited to support more efficient ONN engines via a fine-grained structured pruning technique. A robustness-aware learning method is adopted to guarantee the variation-tolerance of our ONN. Simulation and experimental results show that the proposed ONN achieves one-order-of-magnitude improvement in compactness and efficiency over previous designs with high fidelity and robustness.

Optical neuromorphic computing has demonstrated promising performance with ultra-high computation speed, high bandwidth, and low energy consumption. The traditional optical neural network (ONN) architectures realize neuromorphic computing via electrical weight encoding. However, previous ONN design methodologies can only handle static linear projection with stationary synaptic weights, thus fail to support efficient and flexible computing when both operands are dynamically-encoded light signals. In this work, we propose a novel ONN engine O2NN based on wavelength-division multiplexing and differential detection to enable high-performance, robust, and versatile photonic neural computing with both light operands. Balanced optical weights and augmented quantization are introduced to enhance the representability and efficiency of our architecture. Static and dynamic variations are discussed in detail with a knowledge-distillation-based solution given for robustness improvement. Discussions on hardware cost and efficiency are provided for a comprehensive comparison with prior work. Simulation and experimental results show that the proposed ONN architecture provides flexible, efficient, and robust support for high-performance photonic neural computing with fully-optical operands under low-bit quantization and practical variations.

Deep neural networks (DNN) have shown superior performance in a variety of tasks. As they rapidly evolve, their escalating computation and memory demands make it challenging to deploy them on resource-constrained edge devices. Though extensive efficient accelerator designs, from traditional electronics to emerging photonics, have been successfully demonstrated, they are still bottlenecked by expensive memory accesses due to tremendous gaps between the bandwidth/power/latency of electrical memory and computing cores. Previous solutions fail to fully-leverage the ultra-fast computational speed of emerging DNN accelerators to break through the critical memory bound. In this work, we propose a general and unified framework to trade expensive memory transactions with ultra-fast on-chip computations, directly translating to performance improvement. We are the first to jointly explore the intrinsic correlations and bit-level redundancy within DNN kernels and propose a multi-level in situ generation mechanism with mixed-precision bases to achieve on-the-fly recovery of high-resolution parameters with minimum hardware overhead. Extensive experiments demonstrate that our proposed joint method can boost the memory efficiency by 10-20× with comparable accuracy over four state-of-the-art designs, when benchmarked on ResNet-18/DenseNet-121/MobileNetV2/V3 with various tasks.

Silicon-photonics-based optical neural network (ONN) is a promising hardware platform that could represent a paradigm shift in efficient AI with its CMOS-compatibility, flexibility, ultra-low execution latency, and high energy efficiency. In-situ training on the online programmable photonic chips is appealing but still encounters challenging issues in on-chip implementability, scalability, and efficiency. In this work, we propose a closed-loop ONN on-chip learning framework L2ight to enable scalable ONN mapping and efficient in-situ learning. L2ight adopts a three-stage learning flow that first calibrates the complicated photonic circuit states under challenging physical constraints, then performs photonic core mapping via combined analytical solving and zeroth-order optimization. A subspace learning procedure with multi-level sparsity is integrated into L2ight to enable in-situ gradient evaluation and fast adaptation, unleashing the power of optics for real on-chip intelligence. Extensive experiments demonstrate our proposed L2ight outperforms prior ONN training protocols with 3-order-of-magnitude higher scalability and over 30x better efficiency, when benchmarked on various models and learning tasks. This synergistic framework is the first scalable on-chip learning solution that pushes this emerging field from intractable to scalable and further to efficient for next-generation self-learnable photonic neural chips. From a co-design perspective, L2ight also provides essential insights for hardware-restricted unitary subspace optimization and efficient sparse training. We open-source our framework at the link.

As a promising neuromorphic framework, the optical neural network (ONN) demonstrates ultrahigh inference speed with low energy consumption. However, the previous ONN architectures have high area overhead which limits their practicality. In this article, we propose an area-efficient ONN architecture based on structured neural networks, leveraging optical fast Fourier transform for efficient computation. A two-phase software training flow with structured pruning is proposed to further reduce the optical component utilization. Experimental results demonstrate that the proposed architecture can achieve 2.2- 3.7× area cost improvement compared with the previous singular value decomposition-based architecture with comparable inference accuracy. A novel optical microdisk-based convolutional neural network architecture with joint learnability is proposed as an extension to move beyond Fourier transform and multilayer perception, enabling hardware-aware ONN design space exploration with lower area cost, higher power efficiency, and better noise-robustness.

Hybrid analog/digital precoding in millimeter-wave (mmWave) multi-input multi-ouput (MIMO) systems is capable of achieving the near-optimal full-digital performance at reduced hardware cost and power consumption compared to its full-RF digital counterpart. However, having numerous phase shifters is still costly, especially when the phase shifters are of high resolution. In this paper, we propose a novel twin-resolution phase-shifter network for mmWave MIMO systems, which reduces the power consumption of an entirely high-resolution network, whilst mitigating the severe array gain reduction of an entirely low-resolution network. The connections between the twin phase shifters having different resolutions and the antennas are either fixed or dynamically configured. In the latter, we jointly design the phase-shifter network and the hybrid precoding matrix, where the phase of each entry in the analog precoding matrix can be dynamically designed according to the required resolution. This method is slightly modified for the fixed network's hybrid precoding matrix. Furthermore, we extend the proposed method to multi-user MIMO systems and provide its performance analysis. Our simulation results show that the proposed dynamic hybrid precoding method strikes an attractive performance vs. power consumption trade-off.

As a promising neuromorphic framework, the optical neural network (ONN) demonstrates ultra-high inference speed with low energy consumption. However, the previous ONN architectures have high area overhead which limits their practicality. In this paper, we propose an area-efficient ONN architecture based on structured neural networks, leveraging optical fast Fourier transform for efficient computation. A two-phase software training flow with structured pruning is proposed to further reduce the optical component utilization. Experimental results demonstrate that the proposed architecture can achieve 2.2~3.7× area cost improvement compared with the previous singular value decomposition-based architecture with comparable inference accuracy.

Optical neural networks (ONNs) have attracted extensive attention due to its ultra-high execution speed and low energy consumption. The traditional software-based ONN training, however, suffers the problems of expensive hardware mapping and inaccurate variation modeling while the current on-chip training methods fail to leverage the self-learning capability of ONNs due to algorithmic inefficiency and poor variation- robustness. In this work, we propose an on-chip learning method to resolve the aforementioned problems that impede ONNs' full potential for ultra-fast forward acceleration. We directly optimize optical components using stochastic zeroth-order optimization on-chip, avoiding the traditional high-overhead back-propagation, matrix decomposition, or in situ devicelevel intensity measurements. Experimental results demonstrate that the proposed on-chip learning framework provides an efficient solution to train integrated ONNs with 3~4× fewer ONN forward, higher inference accuracy, and better variation-robustness than previous works.

Optical neural networks (ONNs) demonstrate orders-of-magnitude higher speed in deep learning acceleration than their electronic counterparts. However, limited control precision and device variations induce accuracy degradation in practical ONN implementations. To tackle this issue, we propose a quantization scheme that adapts a full-precision ONN to low-resolution voltage controls. Moreover, we propose a protective regularization technique that dynamically penalizes quantized weights based on their estimated noise-robustness, leading to an improvement in noise robustness. Experimental results show that the proposed scheme effectively adapts ONNs to limited-precision controls and device variations. The resultant four-layer ONN demonstrates higher inference accuracy with lower variances than baseline methods under various control precisions and device noises.

Photonic integrated circuit (PIC), as a promising alternative to traditional CMOS circuit, has demonstrated the potential to accomplish on-chip optical interconnects and computations in ultra-high speed and/or low power consumption. Wavelength division multiplexing (WDM) is widely used in optical communication for enabling multiple signals being processed and transferred independently. In this work, we apply WDM to optical logic PIC synthesis to reduce the PIC area.

Photonic integrated circuit (PIC), as a promising alternative to traditional CMOS circuit, has demonstrated the potential to accomplish on-chip optical signal transmission and computations in ultra-high speed and/or low power consumption. One of the critical challenges of PIC, however, is that its scalability and robustness are limited by cascaded optical power loss and noise error. In this paper, we analyze the scalability and noise robustness challenges facing photonic integrated circuits, for two representative PIC applications: logic computing and neural networks. Automated design algorithms and learning methodologies are proposed to resolve these issues.