Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

All-optical computing with beyond 100-GHz clock rates

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper demonstrates an all-optical recurrent neural network that performs linear operations, nonlinear activations, and memory entirely in the optical domain at clock rates above 100 GHz, more than an order of magnitude faster than…

desk verdict A real all-optical RNN with a credible nonlinear advantage at 10–50 GHz, but the >100 GHz headline rests on one unquantified point and should be tempered. read the letter →

arxiv 2501.05756 v2 pith:MFDQQI3Q submitted 2025-01-10 physics.optics

Gordon H.Y. Li , Midya Parto , Jinhao Ge , Qing-Xin Ji , Maodong Gao , Yan Yu , James Williams , Robert M. Gray
show 4 more authors
Christian R. Leefmans Nicolas Englebert Kerry J. Vahala Alireza Marandi
This is my paper · ORCID
classification physics.optics PACS 42.65.-k42.79.Ta
keywords all-opticalcomputingrecurrentneuralnetworkultrafastnonlinearopticsopticalfrequencycombsperiodicallypoledlithiumniobateneuromorphicphotonicstime-multiplexedphotonicnetworksquantumnoiseimagegeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Modern computer clock rates have stalled near 5 GHz for two decades, and the paper's premise is that this ceiling comes from electronics' response times, not from computation itself. It proposes and experimentally demonstrates an all-optical recurrent neural network (AO-RNN) in which every operation, linear mixing, nonlinear activation, and the recurrent memory, is performed by light, with the clock rate set by the repetition rate of an optical frequency comb. The central experimental result is waveform classification at clock rates from 10 GHz (97.5% accuracy) to 120 GHz (still well above random guessing), exceeding any electronic CPU. The same machine classifies bipartite soliton states from an integrated microcomb with 95.6% accuracy in under 100 ns, forecasts time series at 10 GHz, and generates handwritten-digit-like images seeded purely by quantum fluctuations. The authors identify the finite phase-matching bandwidth of the PPLN nonlinearity and the difficulty of electro-optically generating clean ultrafast inputs as the current limits on accuracy at the highest clock rates.

What carries the argument

The load-bearing object is the time-multiplexed all-optical recurrent neural network: a single active cavity in which each clock tick is one laser pulse, so the clock rate equals the comb repetition rate. The cavity holds the hidden state; intra-cavity delay lines and a Mach-Zehnder interferometer implement weighted linear combinations; the PPLN waveguide implements the element-wise nonlinear activation; and a second interferometer forms the output layer. Training is performed by a model-agnostic forward-only perturbation scheme that estimates the loss gradient from two optical forward passes, then updates modulator voltages on a digital computer, so inference itself remains all-optical.

What would settle it

Directly measure the PPLN's input-output transfer curve with 100-200 GHz pulse trains; if the curve becomes linear, or if classification accuracy at 120 GHz does not improve when the PPLN is replaced by an activation crystal with much larger phase-matching bandwidth, the claim that useful nonlinear all-optical computation survives beyond 100 GHz is falsified. A second check is to repeat the 100 GHz classification using natively optical inputs from a 100-plus GHz microcomb instead of electro-optically time-interleaved waveforms; if accuracy stays near the 10 GHz level, the bottleneck was input generation rather than the computer.

Watch

Extended reading notes

Core claim

The discovery is that a recurrent neural network can be built end-to-end in optics such that the minimum time between successive operations is the pulse period of a laser comb, not an electronic clock. Information is encoded in the coherent amplitude of ultrashort pulses; an active fiber cavity holds the recurrent state, a two-arm Mach-Zehnder interferometer with electro-optic modulators supplies trainable linear couplings, and a reverse-proton-exchange periodically poled lithium niobate (PPLN) waveguide provides a sigmoid-like nonlinear activation via pump-depleted second-harmonic generation. Because the chi(2) nonlinearity responds nearly instantaneously, the authors argue, the nonlinearity does not set a slow timescale; the measured limits come from the PPLN's roughly 100 GHz phase-matching bandwidth, pulse overlap at about 200 GHz, and input generation fidelity. On this architecture the paper claims the first end-to-end optical computer with over 100 GHz clock rates, demonstrated by tasks where inference runs without any electronic conversion of the data stream.

Load-bearing premise

The useful computing claim above 100 GHz rests on the PPLN nonlinear activation still acting as a genuine nonlinearity at those rates; if it merely passes light through in a linear way, the machine reduces to a linear recurrent network, and the paper's own accuracy numbers at 100 GHz (58%, below the 62% linear baseline) show the risk.

Editorial extensions

If this is right

  • An all-optical computer can execute a full recurrent step, linear mixing, nonlinear activation, and memory update, on every clock tick of a 100-plus GHz pulse train, something no electronic CPU does today.
  • For tasks whose inputs are already optical, such as microcomb soliton waveforms, the computer yields 95.6% classification accuracy with less than 100 ns of processing time, making in-situ ultrafast diagnostics feasible.
  • Time-series forecasting runs at 10 GHz in these experiments; the stated limit is photodetector bandwidth (about 25 GHz), not the optical computer, so optical-to-optical use could go faster.
  • The recurrent cavity can be trained to map quantum noise from spontaneous emission to a target image distribution, producing handwritten-digit images without any optical input signal.
  • At clock rates of 10, 50, and 100 GHz, classification accuracy is 97.5%, 92%, and 58%, respectively; the nonlinear activation is responsible for most of the performance, since a linear model at 10 GHz reaches only 62%.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's data imply that its own "nonlinear optical advantage" criterion is not met at the headline 100-plus GHz rates: at 100 GHz the 58% accuracy falls below the 62% linear-model baseline, so the advantage is demonstrated below that clock rate while the over-100 GHz claim stands for operating speed, not for beating linear models on this task.
  • If the accuracy drop from 10 to 100 GHz is dominated by the PPLN phase-matching bandwidth, swapping in a broader-bandwidth activation crystal should recover most of the 97.5% accuracy at 100 GHz; if it does not, the bottleneck is input generation or pulse overlap.
  • The architecture's natural near-term niche is native optical input, since the paper's own electro-optic time-interleaving introduces jitter and phase noise that it names as a fidelity limiter; one testable extension is to run the classifier directly on soliton microcombs with 100-plus GHz repetition rates.
  • The quantum-noise image generation hints that the same cavity could act as a programmable noise-driven sampler for other target distributions, though the paper states that its simple model lacks the expressiveness to handle multi-class image datasets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes and experimentally demonstrates an all-optical recurrent neural network (AO-RNN) built from fiber-optic components, in which linear operations, nonlinear activation, and memory are implemented optically via an active cavity, a Mach-Zehnder interferometer with electro-optic weights, and a PPLN-based nonlinearity. The authors report classification of noisy waveforms at clock rates from 250 MHz to 120 GHz, classification of microcomb soliton states at about 20 GHz, time-series forecasting at 10 GHz, and MNIST image generation seeded from quantum noise. The central claim is that the AO-RNN achieves an all-optical computing clock rate above 100 GHz, exceeding electronic CPUs.

Significance. If the central claim is established, the work would be a notable advance in ultrafast optical computing: an end-to-end optical recurrent network with a nonlinear activation and optical memory, demonstrated on a prototypical temporal classification task and on native ultrafast optical inputs. The 97.5% accuracy at fc=10 GHz, the 95.6% accuracy for soliton-state classification, and the quantum-noise-seeded image generation are interesting proof-of-concept results. The training approach (forward-only, model-agnostic) is also a constructive methodological contribution. However, the headline 'beyond 100-GHz' claim is not yet supported by the quantitative evidence provided, and the paper's own statements about bandwidth and input-generation limitations leave the interpretation of the high-rate measurements unresolved. The significance would be high if the high-rate nonlinear advantage were demonstrated with proper controls and uncertainties; as it stands, the paper mainly demonstrates a credible low-to-mid-GHz all-optical recurrent computer.

major comments (4)
  1. [Results, 'Noisy waveform classification' (Fig. 3b)] The central claim of computing with clock rates above 100 GHz rests on a single unquantified data point: the text states that at fc=120 GHz the accuracy is 'significantly higher than random guessing' but gives no numerical value, no error bar, and no statistical test. Moreover, at fc=100 GHz the reported accuracy (58%) is below the 62% accuracy of the linear model measured at fc=10 GHz, which the authors themselves use to define a 'nonlinear optical advantage.' This undercuts the claim that the nonlinear optical computation retains an advantage at or above 100 GHz. The authors should provide a quantified 120 GHz accuracy with uncertainty, and ideally a linear-model control at the same high clock rate to separate the contribution of the nonlinearity from that of the linear recurrent dynamics.
  2. [Results, 'Noisy waveform classification' (Fig. 3b)] No error bars, confidence intervals, or repeated-trial information are reported for any of the classification accuracies (97.5%, 92%, 58%, and the unquantified 120 GHz point). Because the 100 GHz result is close to the linear baseline and the 120 GHz result is unquantified, it is impossible to assess whether the degradation is statistically significant or whether the 'nonlinear optical advantage' boundary is properly drawn. The authors should report the number of independent runs and the dispersion of results, or otherwise provide statistical significance measures.
  3. [Results, 'Noisy waveform classification' and 'Native ultrafast optical signals'] The paper identifies two factors that could explain the accuracy drop at high fc: the PPLN phase-matching bandwidth of about 100 GHz, after which the nonlinear activation 'begins to degrade,' and the increasing difficulty of electro-optic time-interleaved input generation at higher fc. These two confounds are not disentangled in the measurements. A skeptical reader cannot distinguish between the AO-RNN's own high-rate limitations and corruption of the input waveforms. The authors should either characterize the fidelity of the input signals at each fc (e.g., measure the actual time-interleaved waveforms) or use a native optical input at high rate, as they do for the soliton-classification task, to demonstrate that the computation, rather than the input generation, is responsible for the observed behavior.
  4. [Discussion and Conclusion] The conclusion states that the AO-RNN 'can achieve clock rates > 100 GHz' and 'more than an order-of-magnitude improvement over current digital computers.' Given that the only >100 GHz point is unquantified and the quantified 100 GHz point falls below the linear-model benchmark, this statement is stronger than the evidence presented. The authors should either provide the missing high-rate data and controls, or revise the conclusion to accurately reflect that the demonstrated high-accuracy operation is at 10-50 GHz with degradation at higher rates.
minor comments (5)
  1. [Figure 3 caption] The caption contains a typo: 'seqeunces' should read 'sequences'.
  2. [Materials and Methods, 'CPU clock rates'] The text says 'commericially-available'; this should be spelled 'commercially-available'.
  3. [Materials and Methods, 'Training procedure'] The phrase 'model paramters' should read 'model parameters'.
  4. [Supplementary Information, Section IV] There are unresolved citation placeholders ' [ ? ] ' for the Si3N4 platform and the QPSK modulation; these should be replaced with proper references.
  5. [Data and Code Availability] The statements say data and code are 'available from the corresponding author upon reasonable request' but provide no repository links. Given the paper's reliance on specific measurements and the forward-only training procedure, making the code and processed data publicly available would substantially strengthen reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the AO-RNN's central results are forward experimental measurements with an independent linear-model control; the >100 GHz concern is about evidence strength, not circularity.

full rationale

I walked the claimed derivation chain: the AO-RNN weights are trained by a forward-only perturbation-based training procedure on training samples, then frozen, and the reported accuracies are measured on separate test waveforms (Table I gives a testing size of 800 for noisy waveform classification). The accuracy-versus-clock-rate data in Fig. 3b are therefore genuine forward predictions, not fitted parameters renamed as predictions. The purely linear model is a separate physical measurement with the PPLN bypassed at fc = 10 GHz: 'Repeating the same task using a purely linear model (i.e. without nonlinear activation function) with fc = 10 GHz achieves a classification accuracy of only 62%', so the 'nonlinear optical advantage' criterion is an externally defined benchmark rather than a quantity constructed from the fitted parameters. The self-citations are not load-bearing in a circular sense: Ref. [35] (same group) characterizes the PPLN sigmoid-like activation used as a physical component, and Ref. [11] is invoked only to suggest future thin-film lithium niobate improvements ('we previously demonstrated higher-performance nonlinear activation functions using thin-film lithium niobate (TFLN) with maximum allowable clock rates > 13 THz [11]'); neither citation is used to forbid alternatives or to define the central claim. The manuscript's own stated limitations—the finite PPLN phase-matching bandwidth of about 100 GHz, the increasing difficulty of electro-optic time-interleaved input generation as fc increases, and the general decrease of classification accuracy with fc—are explicit acknowledgments of degraded experimental conditions, not admissions that a result was assumed into existence. The legitimate weakness in the paper—that at fc = 120 GHz only 'significantly higher than random guessing' is quoted, and at fc = 100 GHz accuracy (58%) falls below the 10 GHz linear baseline (62%)—undermines the strength of the >100 GHz nonlinear-advantage claim, but it is an evidentiary and correctness concern, not a circular reduction. No equation in the paper defines a predicted quantity as equal to its fitted input, and no assertion imports a uniqueness result from the authors' prior work. Accordingly, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claims rest primarily on experimentally assembled components and learned parameters rather than on derivation from axioms. The most consequential assumptions are the fidelity of ultrafast input generation and the effectiveness of the PPLN nonlinearity above its phase-matching bandwidth.

free parameters (4)
  • Trained recurrent and output weights (modulator bias and RF voltages) = Not fully specified; V_pi ~ 6 V, perturbation step delta = 0.02 V, learning rate eta = 2e-3
    Learned on a digital computer via forward-only perturbation for each task; central to task performance but not to the clock-rate claim.
  • Hyperparameters: input scaling epsilon, temporal delays T0, T1, T2, output thresholds = epsilon untrained; T0 hyperparameter; T1 ~ 0.3 ns; T2 ~ 1.6 ns; thresholds Z_q/q
    Chosen by hand per task; affect performance and network topology.
  • Diffusion noise schedule sigma_T for image generation = Linear from sigma_100 = 0.001 to sigma_1 = 0.02
    Ad hoc schedule used to define intermediate training targets for the generative task.
  • Pixel time bin period t' and rescaling for image generation = t' = 0.4 ns, rescale to [-1, 1]
    Manual choices mapping cavity output to MNIST pixels.
assumptions (6)
  • standard math RNNs are Turing-complete (Siegelmann-Sontag), so a recurrent architecture can in principle implement general computation.
    Used in the Introduction to motivate the RNN; not exercised by the specific tasks.
  • domain assumption Pump-depleted second-harmonic generation in the PPLN waveguide produces a sigmoid-like amplitude transfer function.
    Cited to ref 35 from the same group; assumed for the nonlinear activation; no direct transfer function is measured in this paper.
  • domain assumption The cavity delay lines and modulators preserve coherent interference such that linear weights are accurately applied to pulse amplitudes.
    The computation relies on phase-stabilized interference; active Pound-Drever-Hall locking is used, but residual phase error is not quantified.
  • domain assumption The nonlinearity is near-instantaneous with a phase-matching bandwidth of roughly 100 GHz, and pulses shorter than 5 ps avoid cross-talk at over 100 GHz.
    Stated in Results and Discussion; at fc above 100 GHz this assumption partially breaks down and accuracy drops, yet the headline claim depends on it.
  • ad hoc to paper For the image generation task, the continuous cavity field can be discretized into time bins representing pixels, and ASE noise is approximately Gaussian.
    The paper states that 'strictly speaking, the nonlinear laser cavity becomes a continuous-field distribution so the concept of discrete time bins is not well-defined'; discretization is a modeling approximation.
  • domain assumption The digital postprocessing for classification (averaging and thresholding) does not bottleneck the computation.
    The class label is assigned after the whole sequence, so the digital step is slower than fc; acceptable but it means the readout is not fully all-optical.

how reviews work

0 comments
Cite this review

Pith. "Pith review of All-optical computing with beyond 100-GHz clock rates." pith.science (2026). https://pith.science/paper/MFDQQI3Q

@misc{pith2026250105756,
  author       = {Pith},
  title        = {Pith review of: All-optical computing with beyond 100-GHz clock rates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MFDQQI3Q}},
  note         = {Machine review of arXiv:2501.05756}
}
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.

Figures

Figures reproduced from arXiv: 2501.05756 by the authors.

Figure 1
Figure 1. FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6 [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Two-optical-cycle pulses from nanophotonic two-color soliton compression

    physics.optics 2025-01 conditional novelty 7.0 of 10

    Nanophotonic lithium niobate waveguides compress 35-fs pulses at 2 µm to 13 fs (two cycles) via quadratic two-color soliton dynamics, experimentally verified by FROG.

  2. Photonic Ising machines toward and beyond a million spins

    physics.optics 2026-07 conditional novelty 4.0 of 10

    Million-spin photonic Ising machines are argued to be within reach via chiplet, free-space, and all-optical spatiotemporal architectures, but only with major engineering advances.

Reference graph

Works this paper leans on

73 extracted references · 65 canonical work pages · cited by 2 Pith papers

  1. [1]

    Choose a random direction vector ∆ ∈{ +δ,−δ}d where d is the number of trainable model parameters, the elements of ∆ are sampled from a Bernoulli distribution ∆ i∼B(1/2) for i = 1, 2,...,d , and δ is the step size

    work page

  2. [2]

    Perturb the model parameters Θ ∈ Rd by ∆ and perform a forward-pass through the model to evaluate the loss functionL(Θ + ∆)

    work page

  3. [3]

    Perturb the model parameters Θ in the opposite direction −∆ and perform a forward-pass through the model to evaluate the loss function L(Θ− ∆)

    work page

  4. [4]

    Estimate the directional derivative of the loss as: ∇∆L(Θ)≈L(Θ + ∆)−L (Θ− ∆) 2∥∆∥ . (3)

    work page

  5. [5]

    For the experimental AO-RNN, the forward-pass steps are performed directly in the optical hardware, but the other training steps are performed on a digital computer

    Update the model parameters: Θ → Θ−η∇∆L(Θ)∆ where η is the learning rate. For the experimental AO-RNN, the forward-pass steps are performed directly in the optical hardware, but the other training steps are performed on a digital computer. During testing, the trained parameters are frozen and so the forward-pass inference is all-optical. The model paramte...

    work page 2024

  6. [6]

    J. L. Hennessy and D. A. Patterson, Computer architecture: a quantitative approach (Elsevier, 2011)

    work page 2011

  7. [7]

    F. L. Bauer, Origins and foundations of computing: in cooperation with Heinz Nixdorf MuseumsForum (Springer Science & Business Media, 2009)

    work page 2009

  8. [8]

    A. W. Burks, Electronic computing circuits of the eniac, Proceedings of the IRE 35, 756 (1947)

    work page 1947

Show all 73 references
  1. [9]

    R. R. Schaller, Moore’s law: past, present and future, IEEE Spectrum 34, 52 (1997)

    1997

  2. [10]

    R. H. Dennard, F. H. Gaensslen, H.-N. Yu, V. L. Rideout, E. Bassous, and A. R. LeBlanc, Design of ion-implanted mosfet’s with very small physical dimensions, IEEE Journal of solid-state circuits 9, 256 (1974)

    1974

  3. [11]

    Backus, Can programming be liberated from the von neumann style? a functional style and its algebra of programs, Communications of the ACM 21, 613 (1978)

    J. Backus, Can programming be liberated from the von neumann style? a functional style and its algebra of programs, Communications of the ACM 21, 613 (1978)

    1978

  4. [12]

    P. L. McMahon, The physics of optical computing, Nature Reviews Physics 5, 717 (2023)

    2023

  5. [13]

    Heinz, J

    R. Heinz, J. Artman, and S. Lee, Matrix multiplication by optical methods, Applied Optics 9, 2161 (1970)

    1970

  6. [14]

    X. Xu, M. Tan, B. Corcoran, J. Wu, A. Boes, T. G. Nguyen, S. T. Chu, B. E. Little, D. G. Hicks, R. Morandotti, et al., 11 tops photonic convolutional accelerator for optical neural networks, Nature 589, 44 (2021)

    2021

  7. [15]

    Feldmann, N

    J. Feldmann, N. Youngblood, M. Karpov, H. Gehring, X. Li, M. Stappers, M. Le Gallo, X. Fu, A. Lukashchuk, A. S. Raja, et al., Parallel convolutional processing using an integrated photonic tensor core, Nature 589, 52 (2021). 21

    2021

  8. [16]

    G. H. Li, R. Sekine, R. Nehra, R. M. Gray, L. Ledezma, Q. Guo, and A. Marandi, All-optical ultrafast relu function for energy-efficient nanophotonic deep learning, Nanophotonics 12, 847 (2023)

    2023

  9. [17]

    Q. Guo, R. Sekine, L. Ledezma, R. Nehra, D. J. Dean, A. Roy, R. M. Gray, S. Jahani, and A. Marandi, Femtojoule femtosecond all-optical switching in lithium niobate nanophotonics, Nature Photonics 16, 625 (2022)

    2022

  10. [18]

    Mourgias-Alexandris, A

    G. Mourgias-Alexandris, A. Tsakyridis, N. Passalis, A. Tefas, K. Vyrsokinos, and N. Pleros, An all-optical neuron with sigmoid activation function, Optics Express 27, 9620 (2019)

    2019

  11. [19]

    Y. Shen, N. C. Harris, S. Skirlo, M. Prabhu, T. Baehr-Jones, M. Hochberg, X. Sun, S. Zhao, H. Larochelle, D. Englund, et al., Deep learning with coherent nanophotonic circuits, Nature Photonics 11, 441 (2017)

    2017

  12. [20]

    Ashtiani, A

    F. Ashtiani, A. J. Geers, and F. Aflatouni, An on-chip photonic deep neural network for image classification, Nature 606, 501 (2022)

    2022

  13. [21]

    Bandyopadhyay, A

    S. Bandyopadhyay, A. Sludds, S. Krastanov, R. Hamerly, N. Harris, D. Bunandar, M. Streshinsky, M. Hochberg, and D. Englund, Single-chip photonic deep neural network with forward-only training, Nature Photonics , 1 (2024)

    2024

  14. [22]

    Feldmann, N

    J. Feldmann, N. Youngblood, C. D. Wright, H. Bhaskaran, and W. H. Pernice, All-optical spiking neurosynaptic networks with self-learning capabilities, Nature 569, 208 (2019)

    2019

  15. [23]

    A. N. Tait, T. F. De Lima, E. Zhou, A. X. Wu, M. A. Nahmias, B. J. Shastri, and P. R. Prucnal, Neuromorphic photonic networks using silicon photonic weight banks, Scientific Reports 7, 7430 (2017)

    2017

  16. [24]

    P. R. Prucnal and B. J. Shastri, Neuromorphic photonics (CRC press, 2017)

    2017

  17. [25]

    Brunner, B

    D. Brunner, B. J. Shastri, M. A. A. Qadasi, H. Ballani, S. Barbay, S. Biasi, P. Bienstman, S. Bilodeau, W. Bogaerts, F. B¨ ohm,et al., Roadmap on neuromorphic photonics, arXiv preprint arXiv:2501.07917 (2025)

    2025 arXiv

  18. [26]

    Marandi, Z

    A. Marandi, Z. Wang, K. Takata, R. L. Byer, and Y. Yamamoto, Network of time-multiplexed optical parametric oscillators as a coherent ising machine, Nature Photonics 8, 937 (2014)

    2014

  19. [27]

    Inagaki, Y

    T. Inagaki, Y. Haribara, K. Igarashi, T. Sonobe, S. Tamate, T. Honjo, A. Marandi, P. L. McMahon, T. Umeki, K. Enbutsu, et al., A coherent ising machine for 2000-node optimization problems, Science 354, 603 (2016)

    2016

  20. [28]

    Honjo, T

    T. Honjo, T. Sonobe, K. Inaba, T. Inagaki, T. Ikuta, Y. Yamada, T. Kazama, K. Enbutsu, T. Umeki, R. Kasahara, et al., 100,000-spin coherent ising machine, Science advances 7, eabh0952 (2021)

    2021

  21. [29]

    Parto, W

    M. Parto, W. Hayenga, A. Marandi, D. N. Christodoulides, and M. Khajavikhan, Realizing spin hamiltonians in nanoscale active photonic lattices, Nature Materials 19, 725 (2020)

    2020

  22. [30]

    X. Lin, Y. Rivenson, N. T. Yardimci, M. Veli, Y. Luo, M. Jarrahi, and A. Ozcan, All-optical machine learning using diffractive deep neural networks, Science 361, 1004 (2018)

    2018

  23. [31]

    Duport, B

    F. Duport, B. Schneider, A. Smerieri, M. Haelterman, and S. Massar, All-optical reservoir computing, Optics Express 20, 22783 (2012). 22

    2012

  24. [32]

    Dejonckheere, F

    A. Dejonckheere, F. Duport, A. Smerieri, L. Fang, J.-L. Oudar, M. Haelterman, and S. Massar, All-optical reservoir computer based on saturation of absorption, Optics express 22, 10868 (2014)

    2014

  25. [33]

    Y. Zuo, B. Li, Y. Zhao, Y. Jiang, Y.-C. Chen, P. Chen, G.-B. Jo, J. Liu, and S. Du, All-optical neural network with nonlinear activation functions, Optica 6, 1132 (2019)

    2019

  26. [34]

    Medsker and L

    L. Medsker and L. C. Jain, Recurrent neural networks: design and applications (CRC press, 1999)

    1999

  27. [35]

    H. T. Siegelmann and E. D. Sontag, On the computational power of neural nets, in Proceedings of the fifth annual workshop on Computational learning theory (1992) pp. 440–449

    1992

  28. [36]

    Leefmans, A

    C. Leefmans, A. Dutt, J. Williams, L. Yuan, M. Parto, F. Nori, S. Fan, and A. Marandi, Topological dissipation in a time-multiplexed photonic resonator network, Nature Physics 18, 442 (2022)

    2022

  29. [37]

    Y. Bai, X. Xu, M. Tan, Y. Sun, Y. Li, J. Wu, R. Morandotti, A. Mitchell, K. Xu, and D. J. Moss, Photonic multiplexing techniques for neuromorphic computing, Nanophotonics 12, 795 (2023)

    2023

  30. [38]

    Fortier and E

    T. Fortier and E. Baumann, 20 years of developments in optical frequency comb technology and applications, Communi- cations Physics 2, 153 (2019)

    2019

  31. [39]

    Langrock and M

    C. Langrock and M. Fejer, Fiber-feedback continuous-wave and synchronously-pumped singly-resonant ring optical para- metric oscillators using reverse-proton-exchanged periodically-poled lithium niobate waveguides, Optics Letters 32, 2263 (2007)

    2007

  32. [40]

    G. H. Li, C. R. Leefmans, J. Williams, R. M. Gray, M. Parto, and A. Marandi, Deep learning with photonic neural cellular automata, Light: Science & Applications 13, 283 (2024)

    2024

  33. [41]

    Z. Yuan, M. Gao, Y. Yu, H. Wang, W. Jin, Q.-X. Ji, A. Feshali, M. Paniccia, J. Bowers, and K. Vahala, Soliton pulse pairs at multiple colours in normal dispersion microresonators, Nature Photonics 17, 977 (2023)

    2023

  34. [42]

    T. J. Kippenberg, A. L. Gaeta, M. Lipson, and M. L. Gorodetsky, Dissipative kerr solitons in optical microresonators, Science 361, eaan8083 (2018)

    2018

  35. [43]

    Herink, F

    G. Herink, F. Kurtz, B. Jalali, D. R. Solli, and C. Ropers, Real-time spectral interferometry probes the internal dynamics of femtosecond soliton molecules, Science 356, 50 (2017)

    2017

  36. [44]

    M. A. Foster, R. Salem, D. F. Geraghty, A. C. Turner-Foster, M. Lipson, and A. L. Gaeta, Silicon-chip-based ultrafast optical oscilloscope, Nature 456, 81 (2008)

    2008

  37. [45]

    Goda and B

    K. Goda and B. Jalali, Dispersive fourier transformation for fast continuous single-shot measurements, Nature Photonics 7, 102 (2013)

    2013

  38. [46]

    D. J. Kane and R. Trebino, Single-shot measurement of the intensity and phase of an arbitrary ultrashort pulse by using frequency-resolved optical gating, Optics Letters 18, 823 (1993)

    1993

  39. [47]

    Yi, Q.-F

    X. Yi, Q.-F. Yang, K. Y. Yang, and K. Vahala, Imaging soliton dynamics in optical microcavities, Nature communications 9, 3565 (2018). 23

    2018

  40. [48]

    R. M. Gray, M. Liu, S. Zhou, A. Roy, L. Ledezma, and A. Marandi, Quadratic-soliton-enhanced mid-ir molecular sensing, Nature Communications 15, 9086 (2024)

    2024

  41. [49]

    Gomber and M

    P. Gomber and M. Haferkorn, High-frequency-trading: High-frequency-trading technologies and their implications for electronic securities trading, Business & Information Systems Engineering 5, 97 (2013)

    2013

  42. [50]

    V. V. Gligorov and M. Williams, Efficient, reliable and fast high-level triggering using a bonsai boosted decision tree, Journal of Instrumentation 8 (02), P02013

  43. [51]

    A. E. Willner, S. Khaleghi, M. R. Chitgarha, and O. F. Yilmaz, All-optical signal processing, Journal of Lightwave Technology 32, 660 (2013)

    2013

  44. [52]

    S. Choi, Y. Salamin, C. Roques-Carmes, R. Dangovski, D. Luo, Z. Chen, M. Horodynski, J. Sloan, S. Z. Uddin, and M. Soljaˇ ci´ c, Photonic probabilistic machine learning using quantum vacuum noise, Nature Communications 15, 7760 (2024)

    2024

  45. [53]

    Roques-Carmes, Y

    C. Roques-Carmes, Y. Salamin, J. Sloan, S. Choi, G. Velez, E. Koskas, N. Rivera, S. E. Kooi, J. D. Joannopoulos, and M. Soljaˇ ci´ c, Biasing the quantum vacuum to control macroscopic probability distributions, Science381, 205 (2023)

    2023

  46. [54]

    J. Ho, A. Jain, and P. Abbeel, Denoising diffusion probabilistic models, Advances in neural information processing systems 33, 6840 (2020)

    2020

  47. [55]

    Lipman, R

    Y. Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, Flow matching for generative modeling, arXiv preprint arXiv:2210.02747 (2022)

    2022 arXiv

  48. [56]

    Deng, The mnist database of handwritten digit images for machine learning research [best of the web], IEEE signal processing magazine 29, 141 (2012)

    L. Deng, The mnist database of handwritten digit images for machine learning research [best of the web], IEEE signal processing magazine 29, 141 (2012)

    2012

  49. [57]

    P. W. Milonni and J. H. Eberly, Laser physics (John Wiley & Sons, 2010)

    2010

  50. [58]

    C. Wang, M. Zhang, X. Chen, M. Bertrand, A. Shams-Ansari, S. Chandrasekhar, P. Winzer, and M. Lonˇ car, Integrated lithium niobate electro-optic modulators operating at cmos-compatible voltages, Nature 562, 101 (2018)

    2018

  51. [59]

    R. M. Gray, T. Zacharias, R. Chawlani, L. Ledezma, R. Sekine, J. A. Williams, and A. Marandi, Soliton pulse compression in lithium niobate nanophotonics, in 2024 Conference on Lasers and Electro-Optics (CLEO) (IEEE, 2024) pp. 1–2

    2024

  52. [60]

    A. Roy, L. Ledezma, L. Costa, R. Gray, R. Sekine, Q. Guo, M. Liu, R. M. Briggs, and A. Marandi, Visible-to-mid-ir tunable frequency comb in nanophotonics, Nature Communications 14, 6549 (2023)

    2023

  53. [61]

    Ledezma, A

    L. Ledezma, A. Roy, L. Costa, R. Sekine, R. Gray, Q. Guo, R. Nehra, R. M. Briggs, and A. Marandi, Octave-spanning tunable infrared parametric oscillators in nanophotonics, Science Advances 9, eadf9711 (2023)

    2023

  54. [62]

    R. M. Gray, R. Sekine, L. Ledezma, G. H. Li, S. Zhou, A. Roy, M. Parto, and A. Marandi, Large-scale time-multiplexed nanophotonic parametric oscillators, arXiv preprint arXiv:2405.17355 (2024)

    2024 arXiv

  55. [63]

    T. Zhou, W. Wu, J. Zhang, S. Yu, and L. Fang, Ultrafast dynamic machine vision with spatiotemporal photonic computing, Science Advances 9, eadg4391 (2023). 24

    2023

  56. [64]

    P. Wang, J. Liang, and L. V. Wang, Single-shot ultrafast imaging attaining 70 trillion frames per second, Nature Commu- nications 11, 2091 (2020)

    2020

  57. [65]

    Maiuri, M

    M. Maiuri, M. Garavelli, and G. Cerullo, Ultrafast spectroscopy: State of the art and open challenges, Journal of the American Chemical Society 142, 3 (2019)

    2019

  58. [66]

    Kikuchi, Fundamentals of coherent optical fiber communications, Journal of Lightwave Technology 34, 157 (2015)

    K. Kikuchi, Fundamentals of coherent optical fiber communications, Journal of Lightwave Technology 34, 157 (2015)

    2015

  59. [67]

    Suh and K

    M.-G. Suh and K. J. Vahala, Soliton microcomb range measurement, Science 359, 884 (2018)

    2018

  60. [68]

    W. C. Swann and N. R. Newbury, Frequency-resolved coherent lidar using a femtosecond fiber laser, Optics Letters 31, 826 (2006)

    2006

  61. [69]

    E. D. Black, An introduction to pound–drever–hall laser frequency stabilization, American Journal of Physics 69, 79 (2001)

    2001

  62. [70]

    Liang, Y

    X. Liang, Y. Zhong, J. Tang, Z. Liu, P. Yao, K. Sun, Q. Zhang, B. Gao, H. Heidari, H. Qian, et al., Rotating neurons for all-analog implementation of cyclic reservoir computing, Nature Communications 13, 1549 (2022)

    2022

  63. [71]

    Standard Performance Evaluation Corporation, Spec benchmarks and tools, https://www.spec.org/benchmarks.html (2024), accessed: 2024-12-31

    2024

  64. [72]

    PassMark Software, Passmark - cpu benchmarks, https://www.cpubenchmark.net/CPU_mega_page.html (2024), accessed: 2024-12-31

    2024

  65. [73]

    All-optical computing with beyond 100-GHz clock rates

    HWBOT, elmor‘s cpu frequency score 9117.75 mhz with core i9 14900ks (8p), https://hwbot.org/submission/5508265_ elmor_cpu_frequency_core_i9_14900ks_(8p)_9117.75_mhz (2024), accessed: 2024-12-31. Supplementary Information for “All-optical computing with beyond 100-GHz clock rat...

    2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.