pith. sign in

arxiv: 2604.10841 · v1 · submitted 2026-04-12 · ⚛️ physics.optics · cs.AI· cs.AR· cs.ET· cs.LG

Harnessing Photonics for Machine Intelligence

Hanqing Zhu , Shupeng Ning , Hongjian Zhou , Ziang Yin , Ray T. Chen , Jiaqi Gu , David Z. Pan This is my paper

Pith reviewed 2026-05-10 15:05 UTC · model grok-4.3

classification ⚛️ physics.optics cs.AIcs.ARcs.ETcs.LG
keywords integrated photonicsphotonic computingAI accelerationco-designdesign automationmachine intelligenceoptical computingelectronic-photonic systems
0
0 comments X

The pith

Integrated photonics can overcome electronic limits in AI by exploiting optical bandwidth and parallelism through cross-layer co-design.

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

The paper reviews how the growing demands of machine intelligence are running into power, memory, and interconnect constraints of conventional electronics. It reframes photonic computing around system-level analysis rather than isolated devices, using a bottleneck-driven taxonomy to map where optical approaches can deliver lasting gains in data movement and computation. The central argument is that cross-layer co-design paired with workload-adaptive programmability, supported by electronic-photonic design automation, is required to move beyond laboratory demonstrations toward scalable, reproducible systems. This matters because it offers a concrete route for hardware to keep pace with evolving AI workloads without relying solely on transistor scaling.

Core claim

Photonics can reshape AI acceleration by leveraging optical bandwidth and parallelism, but only when full-stack electronic-photonic design automation enables closed-loop co-optimization from simulation through physical implementation, allowing sustained efficiency and versatility across application domains.

What carries the argument

Electronic-Photonic Design Automation (EPDA), which performs closed-loop co-optimization across simulation, inverse design, system modeling, and physical implementation.

If this is right

  • Bottleneck-driven taxonomy identifies operating regimes where photonics provides end-to-end sustained benefits over electronics.
  • Workload-adaptive programmability extends versatility as AI application domains continue to evolve.
  • Closed-loop EPDA reduces discrepancies between theoretical designs and fabricated hardware performance.
  • Roadmap supports transition from prototypes to reproducible electronic-photonic ecosystems for machine intelligence.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • EPDA frameworks could enable tighter integration between photonic accelerators and conventional electronic processors in hybrid systems.
  • The taxonomy might be tested by applying it to emerging workloads such as large language model inference to predict efficiency gains.
  • Similar co-design principles could extend to other emerging substrates like neuromorphic or quantum hardware for comparable scaling benefits.

Load-bearing premise

That cross-layer co-design and workload-adaptive programmability can sustain high efficiency and versatility across evolving application domains at scale, moving beyond laboratory prototypes.

What would settle it

A deployed photonic AI accelerator that maintains high efficiency on new workloads without relying on cross-layer co-design or EPDA tools, or a large-scale prototype that fails to deliver promised benefits despite using such methods.

Figures

Figures reproduced from arXiv: 2604.10841 by David Z. Pan, Hanqing Zhu, Hongjian Zhou, Jiaqi Gu, Ray T. Chen, Shupeng Ning, Ziang Yin.

Figure 1
Figure 1. Figure 1: Physical advantages of photonic computing: ultra-low latency via RC-free propagation, high bandwidth density via multiplexing, and superior energy efficiency. • Section IV reviews the EPDA stack required to scale pho￾tonic AI from prototypes to deployable EPICs, including modeling, verification, physical design, and calibration￾aware co-simulation. II. QUANTIFYING PHOTONIC ADVANTAGE: FROM PHYSICAL PROMISE … view at source ↗
Figure 2
Figure 2. Figure 2: System-level energy efficiency vs. compute density comparison between silicon-photonic accelerators and digital GPUs. The axes report effective energy efficiency (TOPS/W) and compute density (TOPS/mm2 ) under reported or derived system-level assumptions. GPU points correspond to NVIDIA V100 [45], A100 [46], H100 SXM [47], and B200 [48], using vendor-reported peak INT8 performance and die area. Photonic poi… view at source ↗
Figure 3
Figure 3. Figure 3: SIMPHONY [33] cross-layer modeling framework for het￾erogeneous electronic-photonic AI systems. A modular PTC can be instantiated from different architecture families under a unified system interface. The framework maps NN operators into GEMM workloads, generates optics-specific dataflow, and evaluates end-to-end metrics using analyzers for memory traffic, area/floorplan, power/energy breakdown, and optica… view at source ↗
Figure 4
Figure 4. Figure 4: (a) Energy efficiency and compute density comparison between photonic and digital electronic hardware. (b) End-to-end energy breakdown comparison across 3 representative PTC families under dynamic attention and static linear workloads with matched computations. All results are simulated using SIMPHONY [33]. Arch Setting: Total of 8 8×8 PTCs with 12 wavelengths, 8-bit precision, and a 5 GHz clock rate. Work… view at source ↗
Figure 5
Figure 5. Figure 5: summarizes the simulated system-level energy breakdowns and energy efficiency trends across three PTC families, using the attention workload as a unified baseline. The results reveal distinct scaling behaviors across these dimensions: Crossbar MZI mesh MRR WB Crossbar MZI mesh MRR WB Crossbar MZI mesh MRR WB Crossbar MZI mesh MRR WB Crossbar MZI mesh MRR WB Crossbar MZI mesh MRR WB 0 5 10 15 20 25 30 2 8 1… view at source ↗
Figure 7
Figure 7. Figure 7: Inverse-designed photonic components and circuit modules can achieve similar functionalities with orders-of-magnitude smaller spatial footprint compared to manual counterparts. (a) Manual [131] and inverse-designed four-mode mode-division multiplexer [132]. (b) Manual [133] and inverse-designed photonic tensor core circuit [129] [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Challenges of photonic inverse design. C. EPDA: Component-level Inverse Design 1) Limitations of manually designed devices Conventional photonic device design largely follows a forward-design workflow: starting from canonical topolo￾gies (e.g., couplers, rings) and tuning a small set of ge￾ometric parameters via simulation sweeps. While effec￾tive for standard building blocks, it becomes a bottleneck when … view at source ↗
Figure 9
Figure 9. Figure 9: Representative PIC layout automation research milestones and trends. circuits, focus shifts to circuit-scale automation that outputs manufacturable GDSII. LiDAR/LiDAR2.0 [174, 175] advances detailed routing via dynamic crossing insertion and curvilinear routing, producing near-DRV-free final layout on WRONoC and photonic-computing designs. In parallel, the work [176] proposes an optical routing flow target… view at source ↗
read the original abstract

The exponential growth of machine-intelligence workloads is colliding with the power, memory, and interconnect limits of the post-Moore era, motivating compute substrates that scale beyond transistor density alone. Integrated photonics is emerging as a candidate for artificial intelligence (AI) acceleration by exploiting optical bandwidth and parallelism to reshape data movement and computation. This review reframes photonic computing from a circuits-and-systems perspective, moving beyond building-block progress toward cross-layer system analysis and full-stack design automation. We synthesize recent advances through a bottleneck-driven taxonomy that delineates the operating regimes and scaling trends where photonics can deliver end-to-end sustained benefits. A central theme is cross-layer co-design and workload-adaptive programmability to sustain high efficiency and versatility across evolving application domains at scale. We further argue that Electronic-Photonic Design Automation (EPDA) will be pivotal, enabling closed-loop co-optimization across simulation, inverse design, system modeling, and physical implementation. By charting a roadmap from laboratory prototypes to scalable, reproducible electronic-photonic ecosystems, this review aims to guide the CAS community toward an automated, system-centric era of photonic machine intelligence.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper is a review synthesizing advances in integrated photonics for AI acceleration. It reframes the field from a circuits-and-systems viewpoint, introduces a bottleneck-driven taxonomy of operating regimes and scaling trends where photonics can provide end-to-end benefits, stresses cross-layer co-design together with workload-adaptive programmability, and argues that Electronic-Photonic Design Automation (EPDA) is essential for closed-loop co-optimization across simulation, inverse design, system modeling, and physical implementation, culminating in a roadmap from laboratory prototypes to scalable electronic-photonic ecosystems.

Significance. If the taxonomy and roadmap hold, the review could help consolidate the photonic-computing literature and steer the community toward system-level, automated design practices that move beyond component-level demonstrations. The explicit synthesis of prior work and the forward proposal for EPDA as an enabling infrastructure are constructive contributions that highlight reproducibility and full-stack considerations.

major comments (2)
  1. [abstract and cross-layer co-design discussion] The central claim that cross-layer co-design and workload-adaptive programmability can sustain high efficiency and versatility across evolving domains at scale (abstract and the section on cross-layer co-design) is load-bearing for the proposed roadmap yet remains largely aspirational; the manuscript does not supply concrete quantitative projections, trade-off analyses, or references to existing photonic-system benchmarks that would demonstrate how these principles overcome current integration and programmability limits.
  2. [EPDA and roadmap section] The assertion that EPDA will be pivotal for closed-loop co-optimization (abstract and the EPDA/roadmap section) is presented without a detailed gap analysis of existing EPDA tools or preliminary case studies showing how simulation-to-physical feedback loops have been or could be realized in photonic AI hardware; this weakens the concreteness of the scalability argument.
minor comments (2)
  1. [taxonomy section] The bottleneck-driven taxonomy would be clearer if each regime were accompanied by explicit quantitative thresholds (e.g., bandwidth, power, or latency targets) drawn from the cited literature.
  2. [roadmap discussion] A few forward-looking statements on versatility could be tempered by brief acknowledgment of documented challenges in photonic integration, such as fabrication variability or thermal sensitivity, to maintain balance.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and insightful comments, which help us strengthen the manuscript's arguments on cross-layer co-design and EPDA. We address each major comment point by point below and outline targeted revisions.

read point-by-point responses

  1. Referee: [abstract and cross-layer co-design discussion] The central claim that cross-layer co-design and workload-adaptive programmability can sustain high efficiency and versatility across evolving domains at scale (abstract and the section on cross-layer co-design) is load-bearing for the proposed roadmap yet remains largely aspirational; the manuscript does not supply concrete quantitative projections, trade-off analyses, or references to existing photonic-system benchmarks that would demonstrate how these principles overcome current integration and programmability limits.

    Authors: We appreciate this observation and agree that additional specificity would enhance the discussion. As a review, the manuscript synthesizes existing literature rather than introducing new data; however, we will revise the cross-layer co-design section to incorporate explicit references to quantitative benchmarks from recent photonic AI accelerators and optical neural network implementations. This will include trade-off analyses drawn from the literature that illustrate efficiency and versatility gains achieved through workload-adaptive programmability and co-design, directly addressing integration and programmability challenges. revision: yes

  2. Referee: [EPDA and roadmap section] The assertion that EPDA will be pivotal for closed-loop co-optimization (abstract and the EPDA/roadmap section) is presented without a detailed gap analysis of existing EPDA tools or preliminary case studies showing how simulation-to-physical feedback loops have been or could be realized in photonic AI hardware; this weakens the concreteness of the scalability argument.

    Authors: We acknowledge the validity of this point. The manuscript positions EPDA as essential but does not provide an in-depth gap analysis. In revision, we will expand the EPDA and roadmap section with a concise review of limitations in current electronic-photonic design tools, supported by references to the literature on inverse design and system-level simulation. We will also include preliminary case studies from photonic hardware demonstrating closed-loop feedback approaches, thereby making the scalability argument more concrete while remaining within the scope of a review. revision: yes

Circularity Check

0 steps flagged

No significant circularity: review synthesizes external literature without internal derivations

full rationale

This is a review and roadmap paper. It contains no original equations, derivations, fitted parameters, or predictions that reduce to the paper's own inputs by construction. All claims are positioned as synthesis of cited prior work or as proposed future directions (e.g., EPDA enabling closed-loop co-optimization). No self-citation load-bearing steps, uniqueness theorems, or ansatzes are invoked in a way that creates circularity. The structure is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

As a review paper, the central claims rest on synthesis of prior photonic computing and AI literature rather than new derivations; no free parameters, axioms, or invented entities are introduced in the abstract.

pith-pipeline@v0.9.0 · 5518 in / 1037 out tokens · 77771 ms · 2026-05-10T15:05:12.065888+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

182 extracted references · 182 canonical work pages · 4 internal anchors

  1. [1]

    Deep learning,

    Y . LeCun, Y . Bengioet al., “Deep learning,”nature, vol. 521, no. 7553, pp. 436–444, 2015

    work page 2015

  2. [2]

    Imagenet classi- fication with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskeveret al., “Imagenet classi- fication with deep convolutional neural networks,” in Advances in Neural Information Processing Systems (NIPS), 2012

    work page 2012

  3. [3]

    Scaling Laws for Neural Language Models

    J. Kaplan, S. McCandlishet al., “Scaling laws for neural language models,”arXiv preprint arXiv:2001.08361, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2001

  4. [4]

    Training Compute-Optimal Large Language Models

    J. Hoffmann, S. Borgeaudet al., “Training compute- optimal large language models,”arXiv preprint arXiv:2203.15556, 2022

    work page internal anchor Pith review arXiv 2022

  5. [5]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wanget al., “Chain-of-thought prompting elicits reasoning in large language models,”Advances in neural information processing systems, vol. 35, pp. 24 824–24 837, 2022

    work page 2022

  6. [6]

    Inference scaling laws: An empirical analysis of compute-optimal inference for llm problem-solving,

    Y . Wu, Z. Sunet al., “Inference scaling laws: An empirical analysis of compute-optimal inference for llm problem-solving,” inThe Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  7. [7]

    Can test-time scaling improve world foundation model?

    W. Cong, H. Zhuet al., “Can test-time scaling improve world foundation model?”arXiv preprint arXiv:2503.24320, 2025

    work page arXiv 2025

  8. [8]

    OpenAI o1 System Card

    A. Jaech, A. Kalaiet al., “Openai o1 system card,”arXiv preprint arXiv:2412.16720, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    D. Guo, D. Yanget al., “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,” arXiv preprint arXiv:2501.12948, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    Cramming more components onto inte- grated circuits,

    G.E. Moore, “Cramming more components onto inte- grated circuits,”Proceedings of the IEEE, vol. 86, no. 1, pp. 82–85, 1998

    work page 1998

  11. [11]

    A 30 year retrospective on dennard’s mos- fet scaling paper,

    M. Bohr, “A 30 year retrospective on dennard’s mos- fet scaling paper,”IEEE Solid-State Circuits Society Newsletter, vol. 12, no. 1, pp. 11–13, 2009

    work page 2009

  12. [12]

    More than moore,

    M.M. Waldrop, “More than moore,”Nature, vol. 530, no. 7589, pp. 144–148, 2016

    work page 2016

  13. [13]

    Towards atomic and close-to- atomic scale manufacturing,

    F. Fang, N. Zhanget al., “Towards atomic and close-to- atomic scale manufacturing,”International Journal of Extreme Manufacturing, vol. 1, no. 1, p. 012001, 2019. V CONCLUSION AND OUTLOOK 17

    work page 2019

  14. [14]

    Computing’s Energy Problem,

    M. Horowitz, “Computing’s Energy Problem,” inISSCC, 2014

    work page 2014

  15. [15]

    Dark silicon and the end of multicore scaling,

    H. Esmaeilzadeh, E. Blemet al., “Dark silicon and the end of multicore scaling,” inProc. ISCA, 2011, pp. 365–376

    work page 2011

  16. [16]

    The future of computing beyond moore’s law,

    J. Shalf, “The future of computing beyond moore’s law,”Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, vol. 378, no. 2166, 2020

    work page 2020

  17. [17]

    Integrating silicon photonics with complementary metal–oxide–semiconductor tech- nologies,

    Y . Wan, W. Heet al., “Integrating silicon photonics with complementary metal–oxide–semiconductor tech- nologies,”Nature Reviews Electrical Engineering, pp. 1–17, 2025

    work page 2025

  18. [18]

    The physics of optical computing,

    P.L. McMahon, “The physics of optical computing,” Nature Reviews Physics, vol. 5, no. 12, pp. 717–734, 2023

    work page 2023

  19. [19]

    Integrated photonic computing beyond the von neumann architecture,

    X.Y . Xu and X.M. Jin, “Integrated photonic computing beyond the von neumann architecture,”ACS Photonics, vol. 10, no. 4, pp. 1027–1036, 2023

    work page 2023

  20. [20]

    Photonic-electronic integrated circuits for high-performance computing and ai acceler- ators,

    S. Ning, H. Zhuet al., “Photonic-electronic integrated circuits for high-performance computing and ai acceler- ators,”Journal of Lightwave Technology, 2024

    work page 2024

  21. [21]

    Are optical transistors the logical next step?

    D.A. Miller, “Are optical transistors the logical next step?”Nature Photonics, vol. 4, no. 1, pp. 3–5, 2010

    work page 2010

  22. [22]

    Analog optical computing,

    D.R. Solli and B. Jalali, “Analog optical computing,” Nature Photonics, vol. 9, no. 11, pp. 704–706, 2015

    work page 2015

  23. [23]

    Universal photonic artificial intelligence acceleration,

    S.R. Ahmed, R. Baghdadiet al., “Universal photonic artificial intelligence acceleration,”Nature, vol. 640, no. 8058, pp. 368–374, 2025

    work page 2025

  24. [24]

    W., and Keutzer, K

    A. Gholami, S. Kimet al., “A survey of quan- tization methods for efficient neural network infer- ence. corr abs/2103.13630 (2021),”arXiv preprint arXiv:2103.13630, 2021

    work page arXiv 2021

  25. [25]

    Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale,

    T. Dettmers, M. Lewiset al., “Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale,”Advances in neural information processing systems, vol. 35, pp. 30 318–30 332, 2022

    work page 2022

  26. [26]

    11 tops photonic convolutional accelerator for optical neural networks,

    X. Xu, M. Tanet al., “11 tops photonic convolutional accelerator for optical neural networks,”Nature, vol. 589, no. 7840, pp. 44–51, 2021

    work page 2021

  27. [27]

    Plasmonic modulator enables <1 fj/bit electro-optic conversion,

    W. Heni, C. Haffneret al., “Plasmonic modulator enables <1 fj/bit electro-optic conversion,”Science, vol. 365, no. 6453, pp. 613–617, 2019

    work page 2019

  28. [28]

    Deep learning with coherent nanophotonic circuits,

    Y . Shen, N.C. Harriset al., “Deep learning with coherent nanophotonic circuits,”Nature photonics, vol. 11, no. 7, pp. 441–446, 2017

    work page 2017

  29. [29]

    Large-scale photonic chiplet taichi empowers 160-tops/w artificial general intelligence,

    Z. Xu, T. Zhouet al., “Large-scale photonic chiplet taichi empowers 160-tops/w artificial general intelligence,” Science, vol. 384, no. 6692, pp. 202–209, 2024

    work page 2024

  30. [30]

    Photonic in-memory computing primitive for spiking neural networks us- ing phase-change materials,

    I. Chakraborty, G. Sahaet al., “Photonic in-memory computing primitive for spiking neural networks us- ing phase-change materials,”Physical Review Applied, vol. 11, no. 1, p. 014063, 2019

    work page 2019

  31. [31]

    All-optical spiking neurosynaptic networks with self-learning capabilities,

    J. Feldmann, N. Youngbloodet al., “All-optical spiking neurosynaptic networks with self-learning capabilities,” Nature, vol. 569, no. 7755, pp. 208–214, 2019

    work page 2019

  32. [32]

    Lightening-transformer: A dynamically-operated optically-interconnected photonic transformer accelerator,

    H. Zhu, J. Guet al., “Lightening-transformer: A dynamically-operated optically-interconnected photonic transformer accelerator,” in2024 IEEE International Symposium on High-Performance Computer Architecture (HPCA). IEEE, 2024, pp. 686–703

    work page 2024

  33. [33]

    Simphony: A device-circuit- architecture cross-layer modeling and simulation frame- work for heterogeneous electronic-photonic ai system,

    Z. Yin, M. Zhanget al., “Simphony: A device-circuit- architecture cross-layer modeling and simulation frame- work for heterogeneous electronic-photonic ai system,” arXiv preprint arXiv:2411.13715, 2024

    work page arXiv 2024

  34. [34]

    Slow and fast light in optical fibres,

    L. Thévenaz, “Slow and fast light in optical fibres,” Nature photonics, vol. 2, no. 8, pp. 474–481, 2008

    work page 2008

  35. [35]

    Pozar,Microwave engineering: theory and tech- niques

    D.M. Pozar,Microwave engineering: theory and tech- niques. John wiley & sons, 2021

    work page 2021

  36. [36]

    Weste and D

    N.H. Weste and D. Harris,CMOS VLSI design: a circuits and systems perspective. Pearson Education India, 2015

    work page 2015

  37. [37]

    Fine-grained dvfs using on-chip regulators,

    S. Eyerman and L. Eeckhout, “Fine-grained dvfs using on-chip regulators,”ACM Transactions on Architecture and Code Optimization (TACO), vol. 8, no. 1, pp. 1–24, 2011

    work page 2011

  38. [38]

    Power consumption in cmos circuits,

    L.L. Ng, K.H. Yeapet al., “Power consumption in cmos circuits,” inElectromagnetic Field in Advancing Science and Technology. IntechOpen, 2022

    work page 2022

  39. [39]

    Griffiths and D.F

    D.J. Griffiths and D.F. Schroeter,Introduction to quan- tum mechanics. Cambridge university press, 2018

    work page 2018

  40. [40]

    Saleh and M.C

    B.E. Saleh and M.C. Teich,Fundamentals of photonics, 2 volume set. john Wiley & sons, 2019

    work page 2019

  41. [41]

    Massively scalable kerr comb-driven silicon photonic link,

    A. Rizzo, A. Novicket al., “Massively scalable kerr comb-driven silicon photonic link,”Nature Photonics, vol. 17, no. 9, pp. 781–790, 2023

    work page 2023

  42. [42]

    Parallel convolu- tional processing using an integrated photonic tensor core,

    J. Feldmann, N. Youngbloodet al., “Parallel convolu- tional processing using an integrated photonic tensor core,”Nature, vol. 589, no. 7840, pp. 52–58, 2021

    work page 2021

  43. [43]

    Space-efficient optical computing with an integrated chip diffractive neural network,

    H. Zhu, J. Zouet al., “Space-efficient optical computing with an integrated chip diffractive neural network,” Nature communications, vol. 13, no. 1, p. 1044, 2022

    work page 2022

  44. [44]

    Integrated photonic meta- system for image classifications at telecommunication wavelength,

    Z. Wang, L. Changet al., “Integrated photonic meta- system for image classifications at telecommunication wavelength,”Nature communications, vol. 13, no. 1, p. 2131, 2022

    work page 2022

  45. [45]

    NVIDIA V100 Tensor Core GPU Datasheet,

    NVIDIA, “NVIDIA V100 Tensor Core GPU Datasheet,” Jan. 2020, uS-1165301- R5, Jan 2020. [Online]. Available: https://images.nvidia.com/content/technologies/volta/ pdf/volta-v100-datasheet-update-us-1165301-r5.pdf

    work page 2020

  46. [46]

    NVIDIA A100 Tensor Core GPU Datasheet,

    NVIDIA, “NVIDIA A100 Tensor Core GPU Datasheet,” May 2022, 2188504, May 2022. [Online]. Available: https://www.nvidia.com/ content/dam/en-zz/Solutions/Data-Center/a100/pdf/ nvidia-a100-datasheet-nvidia-us-2188504-web.pdf

    work page 2022

  47. [47]

    NVIDIA H100 Tensor Core GPU Datasheet,

    NVIDIA, “NVIDIA H100 Tensor Core GPU Datasheet,” Feb. 2023, 2569583, Feb 2023. [Online]. Available: https://www.cisco.com/c/dam/en/us/products/collateral/ servers-unified-computing/ucs-c-series-rack-servers/ nvidia-h100-80-gpu.pdf

    work page 2023

  48. [48]

    PCF Summary for NVIDIA HGX B200 (Datasheet),

    NVIDIA, “PCF Summary for NVIDIA HGX B200 (Datasheet),” Jul. 2025, 4069550, Jul 2025. [Online]. Available: https://images.nvidia.com/aem-dam/Solutions/ documents/HGX-B200-PCF-Summary.pdf

    work page 2025

  49. [49]

    Microring weight banks,

    A.N. Tait, A.X. Wuet al., “Microring weight banks,” IEEE Journal of Selected Topics in Quantum Electronics, V CONCLUSION AND OUTLOOK 18 vol. 22, no. 6, pp. 312–325, 2016

    work page 2016

  50. [50]

    Photonic tensor cores for machine learning,

    M. Miscuglio and V .J. Sorger, “Photonic tensor cores for machine learning,”Applied Physics Reviews, vol. 7, no. 3, p. 031404, 07 2020

    work page 2020

  51. [51]

    A compact butterfly-style silicon photonic–electronic neural chip for hardware-efficient deep learning,

    C. Feng, J. Guet al., “A compact butterfly-style silicon photonic–electronic neural chip for hardware-efficient deep learning,”Acs Photonics, vol. 9, no. 12, pp. 3906– 3916, 2022

    work page 2022

  52. [52]

    Hardware-efficient photonic tensor core: accelerating deep neural networks with structured compression,

    S. Ning, H. Zhuet al., “Hardware-efficient photonic tensor core: accelerating deep neural networks with structured compression,”Optica, vol. 12, no. 7, pp. 1079– 1089, 2025

    work page 2025

  53. [53]

    TeMPO: Efficient time- multiplexed dynamic photonic tensor core for edge AI with compact slow-light electro-optic modulator,

    M. Zhang, D. Yinet al., “TeMPO: Efficient time- multiplexed dynamic photonic tensor core for edge AI with compact slow-light electro-optic modulator,” Journal of Applied Physics, vol. 135, no. 22, p. 223105, 06 2024

    work page 2024

  54. [54]

    SCATTER: Algorithm-Circuit Co-Sparse Photonic Accelerator with Thermal-Tolerant, Power-Efficient In-situ Light Redistribution,

    Z. Yin, N. Gangiet al., “SCATTER: Algorithm-Circuit Co-Sparse Photonic Accelerator with Thermal-Tolerant, Power-Efficient In-situ Light Redistribution,” inProc. IC- CAD, 2024

    work page 2024

  55. [55]

    An electro-photonic system for accelerating deep neural networks,

    C. Demirkiran, F. Eriset al., “An electro-photonic system for accelerating deep neural networks,”J. Emerg. Technol. Comput. Syst., vol. 19, no. 4, Sep. 2023

    work page 2023

  56. [56]

    Squeezelight: A multi-operand ring-based optical neural network with cross-layer scala- bility,

    J. Gu, C. Fenget al., “Squeezelight: A multi-operand ring-based optical neural network with cross-layer scala- bility,”IEEE TCAD, vol. 42, no. 3, pp. 807–819, 2023

    work page 2023

  57. [57]

    Tomfun: A tensorized optical multimodal fusion network,

    X. Xiao, Y . Zhaoet al., “Tomfun: A tensorized optical multimodal fusion network,”APL Machine Learning, vol. 3, no. 1, p. 016121, 03 2025

    work page 2025

  58. [58]

    An efficient general-purpose optical accelerator for neural networks,

    S. Fei, A. Eldebikyet al., “An efficient general-purpose optical accelerator for neural networks,” inProc. ASP- DAC, 2025, p. 1070–1076

    work page 2025

  59. [59]

    Oisa: Architecting an optical in-sensor accelerator for efficient visual computing,

    M. Morsali, S. Tabrizchiet al., “Oisa: Architecting an optical in-sensor accelerator for efficient visual computing,” inProc. DATE, 2024

    work page 2024

  60. [60]

    Ultrafast silicon photonic reservoir computing engine delivering over 200 tops,

    D. Wang, Y . Nieet al., “Ultrafast silicon photonic reservoir computing engine delivering over 200 tops,” Nature Communications, vol. 15, no. 1, p. 10841, Dec 2024

    work page 2024

  61. [61]

    NEOCNN: NTT-Enabled Optical Convolution Neural Network Accelerator,

    X. Li, Y . Liuet al., “NEOCNN: NTT-Enabled Optical Convolution Neural Network Accelerator,” inProc. ICS, 2024, p. 352–362

    work page 2024

  62. [62]

    Highly efficient photonic convolver via lossless mode-division fan-in,

    S. Sun, S. Zhanget al., “Highly efficient photonic convolver via lossless mode-division fan-in,”Nature Communications, vol. 16, no. 1, p. 7513, Aug 2025

    work page 2025

  63. [63]

    Large-scale and energy- efficient tensorized optical neural networks on iii–v-on- silicon moscap platform,

    X. Xiao, M.B. Onet al., “Large-scale and energy- efficient tensorized optical neural networks on iii–v-on- silicon moscap platform,”APL Photonics, vol. 6, no. 12, p. 126107, 12 2021

    work page 2021

  64. [64]

    In-memory photonic dot- product engine with electrically programmable weight banks,

    W. Zhou, B. Donget al., “In-memory photonic dot- product engine with electrically programmable weight banks,”Nature Communications, vol. 14, no. 1, p. 2887, May 2023

    work page 2023

  65. [65]

    Hypermultiplexed integrated photonics–based optical tensor processor,

    S. Ou, K. Xueet al., “Hypermultiplexed integrated photonics–based optical tensor processor,”Science Ad- vances, vol. 11, no. 23, p. eadu0228, 2025

    work page 2025

  66. [66]

    Photonic systolic array for all- optical matrix–matrix multiplication,

    J. Kim, Q. Zhouet al., “Photonic systolic array for all- optical matrix–matrix multiplication,”Laser & Photonics Reviews, vol. n/a, no. n/a, p. e01995, 2025

    work page 2025

  67. [67]

    Integrated multi-operand optical neurons for scalable and hardware-efficient deep learn- ing,

    C. Feng, J. Guet al., “Integrated multi-operand optical neurons for scalable and hardware-efficient deep learn- ing,”Nanophotonics, vol. 13, no. 12, pp. 2193–2206, 2024

    work page 2024

  68. [68]

    Compact optical convolution processing unit based on multimode interference,

    X. Meng, G. Zhanget al., “Compact optical convolution processing unit based on multimode interference,”Nature Communications, vol. 14, no. 1, p. 3000, 2023

    work page 2023

  69. [69]

    Multimodal deep learning using on-chip diffractive optics with in situ training capability,

    J. Cheng, C. Huanget al., “Multimodal deep learning using on-chip diffractive optics with in situ training capability,”Nature Communications, vol. 15, no. 1, p. 6189, 2024

    work page 2024

  70. [70]

    Neuromorphic photonic networks using silicon photonic weight banks,

    A.N. Tait, T.F. De Limaet al., “Neuromorphic photonic networks using silicon photonic weight banks,”Scientific reports, vol. 7, no. 1, p. 7430, 2017

    work page 2017

  71. [71]

    Squeezelight: Towards scalable op- tical neural networks with multi-operand ring resonators,

    J. Gu, C. Fenget al., “Squeezelight: Towards scalable op- tical neural networks with multi-operand ring resonators,” inProc. DATE. IEEE, 2021, pp. 238–243

    work page 2021

  72. [72]

    Microring-based multi-operand optical neurons with on-chip trainable nonlinearity,

    S. Ning, H. Zhuet al., “Microring-based multi-operand optical neurons with on-chip trainable nonlinearity,” inProc. CLEO. Optica Publishing Group, 2025, p. AA120_1

    work page 2025

  73. [73]

    M3icro: Machine learning-enabled compact photonic tensor core based on programmable multi-operand multimode interference,

    J. Gu, H. Zhuet al., “M3icro: Machine learning-enabled compact photonic tensor core based on programmable multi-operand multimode interference,”APL Machine Learning, vol. 2, no. 1, 2024

    work page 2024

  74. [74]

    End-to-end closed-loop optoelec- tronic computing breaking precision–accuracy coupling,

    J. Li, X. Menget al., “End-to-end closed-loop optoelec- tronic computing breaking precision–accuracy coupling,” Advanced Photonics, vol. 8, no. 1, pp. 016 005–016 005, 2026

    work page 2026

  75. [75]

    On-chip wavefront shaping with dielectric metasurface,

    Z. Wang, T. Liet al., “On-chip wavefront shaping with dielectric metasurface,”Nature communications, vol. 10, no. 1, p. 3547, 2019

    work page 2019

  76. [76]

    Diffractive tensorized unit for million-tops general-purpose computing,

    C. Wang, Y . Chenget al., “Diffractive tensorized unit for million-tops general-purpose computing,”Nature Photonics, pp. 1–10, 2025

    work page 2025

  77. [77]

    On-chip reconfigurable diffrac- tive optical neural network based on sb2s3,

    Y . Wang, W. Linet al., “On-chip reconfigurable diffrac- tive optical neural network based on sb2s3,”Optics Express, vol. 33, no. 2, pp. 1810–1826, 2025

    work page 2025

  78. [78]

    Tops-speed complex-valued convolutional accelerator for feature extraction and inference,

    Y . Bai, Y . Xuet al., “Tops-speed complex-valued convolutional accelerator for feature extraction and inference,”Nature Communications, vol. 16, no. 1, p. 292, 2025

    work page 2025

  79. [79]

    High-order tensor flow processing using integrated photonic circuits,

    S. Xu, J. Wanget al., “High-order tensor flow processing using integrated photonic circuits,”Nature communica- tions, vol. 13, no. 1, p. 7970, 2022

    work page 2022

  80. [80]

    Integrated wdm-compatible optical mode division multiplexing neural network accelerator,

    R. Yin, H. Xiaoet al., “Integrated wdm-compatible optical mode division multiplexing neural network accelerator,”Optica, vol. 10, no. 12, pp. 1709–1718, 2023

    work page 2023

Showing first 80 references.