arxiv: 2604.22660 · v1 · submitted 2026-04-24 · ⚛️ physics.optics

Recognition: unknown

Fully multiplexed photonic tensor computing

Aolong Sun , Junhao Zhao , Fangchen Hu , Sizhe Xing , Yuqin Yuan , Jialin He , Yongzhu Hu , Xuyu Deng

show 16 more authors

Yinjun Liu Ouhan Huang Baiheng Zhao Hancheng Liu Tian Dong Jingkai Zhou Haoyang Sun Liang Chen Chao Shen Feng Bao Ziwei Li Jianyang Shi Wei Chu Bowei Dong Nan Chi Junwen Zhang

Authors on Pith no claims yet

Pith reviewed 2026-05-08 10:17 UTC · model grok-4.3

classification ⚛️ physics.optics

keywords photonic tensor computingmultiplexed photonic coresilicon photonicstensor operationsoptical multiplexingAI inferenceparallel computinghyperspectral classification

0 comments

The pith

A photonic tensor core multiplies parallelism by using five dimensions of light in one optical field.

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

The paper seeks to establish that inverse-designed silicon photonics can create a tensor core capable of simultaneously exploiting wavelength, radio-frequency, guided-mode, time, and space multiplexing so that computational parallelism scales multiplicatively rather than additively. A sympathetic reader would care because tensor operations drive most modern AI workloads, and photonic platforms have long promised high speed and low energy but have been limited by the difficulty of controlling all optical degrees of freedom at once. If the uniform computation holds across channels, the design delivers an estimated 69.12 TOPS throughput and supports 1,800 parallel input streams inside a single core while enabling new applications such as hyperspectral classification and parallel fault diagnosis. Experimental benchmarks confirm the approach works for arithmetic operations, image convolution, and digit recognition.

Core claim

FieldCore is a fully multiplexed photonic tensor core that jointly harnesses wavelength, radio-frequency, guided-mode, time and space dimensions, thereby enabling parallelism to scale multiplicatively within a single optical field. Enabled by inverse-designed silicon photonics, FieldCore preserves a uniform programmed computation across all multiplexed channels in parallel. Experimentally, the core validates performance from ultra-high-baudrate arithmetic operations to high-fidelity image convolution and parallel handwritten-digit recognition. It further unlocks high-dimensional hyperspectral classification and massively parallel mechanical fault diagnosis. The architecture supports an estim

What carries the argument

FieldCore, the photonic tensor core that uses inverse-designed silicon photonics to enforce uniform computation across five simultaneous multiplexing dimensions of a single optical field.

If this is right

The core reaches an aggregate throughput of 69.12 tera operations per second.
A single core can process up to 1,800 parallel input streams.
High-dimensional hyperspectral classification becomes feasible inside one optical field.
Massively parallel mechanical fault diagnosis can run concurrently with other tasks.
Ultra-high-baudrate arithmetic and high-fidelity image convolution are demonstrated without separate hardware for each dimension.

Where Pith is reading between the lines

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

The multiplicative scaling could shrink the physical size of photonic accelerators that must handle multi-modal sensor data.
If crosstalk remains low, the same inverse-design method might extend to additional optical dimensions or larger arrays.
Real-time industrial monitoring systems could integrate the parallel fault-diagnosis mode directly with existing silicon-photonic fabrication flows.
Dynamic electronic control of the multiplexed channels might allow the same hardware to switch between different tensor workloads without redesign.

Load-bearing premise

Inverse-designed silicon photonics can maintain uniform programmed computation across wavelength, radio-frequency, guided-mode, time, and space dimensions without significant crosstalk, loss, or degradation between channels.

What would settle it

Direct measurement of substantial crosstalk, increased computation error, or loss of output fidelity when the device is operated with all five multiplexing dimensions active at once.

read the original abstract

Tensor operations dominate modern computational workloads, yet their further acceleration demands hardware platforms with greater parallelism. Although photonic computing provides a compelling route for parallel processing, fully exploiting all native multiplexing dimensions of optical fields is impeded by the challenges in routing and programming light in all dimensions simultaneously. Here we introduce FieldCore, a fully multiplexed photonic tensor core that jointly harnesses wavelength, radio-frequency, guided-mode, time and space dimensions, thereby enabling parallelism to scale multiplicatively within a single optical field. Enabled by inverse-designed silicon photonics, FieldCore preserves a uniform programmed computation across all multiplexed channels in parallel. Experimentally, we validate and benchmark its performance from ultra-high-baudrate arithmetic operations to high-fidelity image convolution and parallel handwritten-digit recognition. We further use FieldCore to unlock applications that naturally require high-dimensional data processing, such as high-dimensional hyperspectral classification and massively parallel mechanical fault diagnosis. Our FieldCore supports an estimated aggregate compute throughput of 69.12 tera operations per second (TOPS) and accommodates up to 1,800 parallel input streams within a single core, establishing a scalable paradigm for fully multiplexed photonic tensor computing and AI inference.

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.

Desk Editor's Note private letter to a colleague

FieldCore shows a single photonic core multiplexing five dimensions for tensor ops with claimed 69 TOPS, but the evidence for uniform fidelity across all channels at once is still thin.

read the letter

The paper's main contribution is demonstrating a silicon photonic tensor core, called FieldCore, that runs wavelength, RF, guided-mode, time, and space multiplexing simultaneously inside one device. Inverse design is used to keep the programmed tensor operation consistent across those channels, which lets parallelism scale multiplicatively rather than additively. That combination in a single core is new compared with earlier photonic accelerators that typically use only two or three dimensions at a time. They show the device on arithmetic at high baud rates, convolution, digit recognition, hyperspectral classification, and fault diagnosis, with headline figures of 69.12 TOPS and 1,800 parallel streams. Those application demos and the device integration are the parts that hold up from the description. The soft spot is the missing quantitative support for the uniformity claim when all five dimensions operate together. The abstract gives no crosstalk numbers, no error bars on the tensor results, and no clear protocol showing full 5-way measurements versus partial subsets. Inverse design works well for fixed conditions, but wavelength-dependent scattering, mode dispersion, or RF phase shifts could introduce non-uniformity that does not cancel across every combination; the stress-test note on inter-dimensional crosstalk is a fair point to check. If the reported tasks were validated only in lower-dimensional regimes, the multiplicative scaling to 1,800 streams rests on an extrapolation whose error is not bounded in the given information. This is for hardware groups working on photonic AI accelerators who want to see multi-dimensional multiplexing tried on real silicon. A reader focused on experimental device design and application mapping would get concrete ideas from the architecture and the task list, even while wanting tighter data on fidelity. I would bring it to a reading group to discuss the multiplexing approach and the inverse-design choices. I would not cite the performance numbers yet. It deserves peer review because the integration is novel and the experiments target relevant workloads; referees can ask for the missing crosstalk and error data without the paper being incoherent on its own terms.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces FieldCore, a photonic tensor core that simultaneously multiplexes across wavelength, radio-frequency, guided-mode, time, and space dimensions using inverse-designed silicon photonics. It claims this enables multiplicative scaling of parallelism within a single optical field, with experimental validation on arithmetic operations, image convolution, handwritten-digit recognition, hyperspectral classification, and mechanical fault diagnosis. The work reports an estimated aggregate throughput of 69.12 TOPS and capacity for up to 1,800 parallel input streams.

Significance. If the uniformity of the programmed tensor operation across all five dimensions is experimentally confirmed without substantial inter-dimensional crosstalk or fidelity loss, the result would establish a new paradigm for scaling photonic compute density multiplicatively rather than additively, with direct implications for high-throughput AI inference and high-dimensional sensing applications.

major comments (2)

[Abstract] Abstract: the claim that inverse-designed silicon photonics 'preserves a uniform programmed computation across all multiplexed channels in parallel' is load-bearing for the multiplicative scaling to 1,800 streams and 69.12 TOPS, yet the abstract provides no quantitative crosstalk, insertion-loss variation, or fidelity metrics measured under simultaneous five-dimensional operation; if the reported experiments used lower-dimensional subsets, the full-regime extrapolation lacks an error bound.
[Abstract] Abstract (experimental validation paragraph): the demonstrations of arithmetic, convolution, recognition, and classification tasks are presented without error bars, measurement protocols, or explicit confirmation that all five dimensions were active concurrently; this leaves open whether the reported performance reflects the claimed joint multiplexing or sequential/partial configurations.

minor comments (1)

The manuscript would benefit from a dedicated methods or supplementary section detailing the inverse-design optimization constraints and the specific figure-of-merit used to ensure uniformity across wavelength, RF, and mode channels.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments highlighting the need for greater specificity in the abstract regarding quantitative metrics and experimental conditions. We address each point below and propose targeted revisions to improve clarity without altering the manuscript's core claims.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that inverse-designed silicon photonics 'preserves a uniform programmed computation across all multiplexed channels in parallel' is load-bearing for the multiplicative scaling to 1,800 streams and 69.12 TOPS, yet the abstract provides no quantitative crosstalk, insertion-loss variation, or fidelity metrics measured under simultaneous five-dimensional operation; if the reported experiments used lower-dimensional subsets, the full-regime extrapolation lacks an error bound.

Authors: We agree that the abstract would be strengthened by including key quantitative metrics to support the uniformity claim under full multiplexing. The main text and supplementary information provide detailed characterizations of crosstalk, insertion-loss variation, and fidelity across the multiplexed dimensions, including under simultaneous operation. To directly address the concern about extrapolation, we will revise the abstract to incorporate representative metrics and an associated error bound for the reported scaling. We will also add a brief clause confirming that the foundational validations used concurrent five-dimensional multiplexing. revision: partial
Referee: [Abstract] Abstract (experimental validation paragraph): the demonstrations of arithmetic, convolution, recognition, and classification tasks are presented without error bars, measurement protocols, or explicit confirmation that all five dimensions were active concurrently; this leaves open whether the reported performance reflects the claimed joint multiplexing or sequential/partial configurations.

Authors: We acknowledge the value of making the abstract more self-contained on this point. The full manuscript details the measurement protocols in the Methods section and includes error bars in the corresponding figures and tables for each task. All reported demonstrations of arithmetic operations, convolution, recognition, and classification were performed with the five dimensions active concurrently, as stated in the experimental sections. We will revise the abstract's experimental validation paragraph to include a concise confirmation of concurrent operation, reference to error bars, and a brief note on protocols. revision: yes

Circularity Check

0 steps flagged

No circularity: experimental device demonstration with measured performance

full rationale

The paper describes an experimental photonic tensor core using inverse-designed silicon photonics for multi-dimensional multiplexing. All reported performance figures (69.12 TOPS, 1800 streams, arithmetic/convolution/classification results) are presented as direct experimental outcomes and estimates from measurements rather than predictions derived from fitted parameters or self-referential equations. No load-bearing derivations, ansatzes, or uniqueness theorems are invoked that reduce to the paper's own inputs by construction. The central claims rest on hardware implementation and validation, which are externally falsifiable via replication of the experiments. No self-citations are load-bearing in the provided text, and no renaming of known results or smuggling of assumptions occurs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The work rests on standard assumptions of silicon photonics and inverse design; no new physical entities are postulated.

free parameters (1)

Channel counts per dimension
Specific numbers of wavelengths, RF carriers, modes, time slots, and spatial channels are chosen to reach the stated 1,800 streams and 69.12 TOPS; exact selection criteria are not detailed in the abstract.

axioms (1)

domain assumption Inverse-designed silicon photonic circuits can maintain uniform computation response across all multiplexed dimensions simultaneously
This premise is required for the claim that computation remains uniform across the five dimensions.

pith-pipeline@v0.9.0 · 5581 in / 1379 out tokens · 46290 ms · 2026-05-08T10:17:44.029246+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

65 extracted references · 8 canonical work pages

[1]

Existing photonic processors still lack a general tensor-computing framework that systematically harnesses the native parallel dimensions of guided optical signals

The challenge is therefore not merely to map data onto additional dimensions, but to preserve the same computation across the full multiplexed signal space with high fidelity22,36. Existing photonic processors still lack a general tensor-computing framework that systematically harnesses the native parallel dimensions of guided optical signals. Here we add...
[2]

GPT-4 Technical Report

OpenAI et al. GPT-4 Technical Report. Preprint at https://doi.org/10.48550/arXiv.2303.08774 (2024)

work page Pith review doi:10.48550/arxiv.2303.08774 2024
[3]

Hong, D. et al. Hyperspectral imaging. Nat Rev Methods Primers 6, 19 (2026)

2026
[4]

Hong, S. et al. Versatile parallel signal processing with a scalable silicon photonic chip. Nat Commun 16, 288 (2025)

2025
[5]

Modha, D. S. et al. Neural inference at the frontier of energy, space, and time. Science 382, 329–335 (2023)

2023
[6]

Silvano, C. et al. A Survey on Deep Learning Hardware Accelerators for Heterogeneous HPC Platforms. ACM Comput. Surv. 57, 1–39 (2025)

2025
[7]

Zhang, Y . et al. Direct tensor processing with coherent light. Nat. Photon. 1–7 (2025) doi:10.1038/s41566-025-01799-7

work page doi:10.1038/s41566-025-01799-7 2025
[8]

Pellauer, M. et al. Symphony: Orchestrating Sparse and Dense Tensors with Hierarchical Heterogeneous Processing. ACM Trans. Comput. Syst. 41, 1–30 (2023)

2023
[9]

Huang, Q., Tsai, P.-A., Emer, J. S. & Parashar, A. Mind the Gap: Attainable Data Movement and Operational Intensity Bounds for Tensor Algorithms. in 2024 ACM/IEEE 51st Annual International Symposium on Computer Architecture (ISCA) 150–166 (2024). doi:10.1109/ISCA59077.2024.00021

work page doi:10.1109/isca59077.2024.00021 2024
[10]

Zhou, H. et al. Photonic matrix multiplication lights up photonic accelerator and beyond. Light Sci Appl 11, 30 (2022)

2022
[11]

Shen, Y . et al. Deep learning with coherent nanophotonic circuits. Nature Photon 11, 441–446 (2017)

2017
[12]

Dong, B. et al. Higher-dimensional processing using a photonic tensor core with continuous-time data. Nat. Photon. 17, 1080–1088 (2023)

2023
[13]

Bente, I. et al. The potential of multidimensional photonic computing. Nat Rev Phys 7, 439–450 (2025)

2025
[14]

Xu, Z. et al. Large-scale photonic chiplet Taichi empowers 160-TOPS/W artificial general intelligence. Science 384, 202– 209 (2024)

2024
[15]

McMahon, P. L. The physics of optical computing. Nat Rev Phys 5, 717–734 (2023)

2023
[16]

Ashtiani, F., Idjadi, M. H. & Kim, K. Integrated photonic neural network with on-chip backpropagation training. Nature 1–6 (2026) doi:10.1038/s41586-026-10262-8

work page doi:10.1038/s41586-026-10262-8 2026
[17]

Theory of Dielectric Optical Waveguides

Marcuse, D. Theory of Dielectric Optical Waveguides. (Academic Press, New York, 1974)

1974
[18]

Subcarrier multiplexing for high- speed optical transmission

Rongqing Hui et al. Subcarrier multiplexing for high- speed optical transmission. J. Lightwave Technol. 20, 417–427 (2002)

2002
[19]

& Zou, W

Xu, S., Wang, J., Yi, S. & Zou, W. High-order tensor flow processing using integrated photonic circuits. Nat Commun 13, 7970 (2022)

2022
[20]

Bai, B. et al. Microcomb-based integrated photonic processing unit. Nat Commun 14, 66 (2023)

2023
[21]

Hu, Y . et al. Integrated lithium niobate photonic computing circuit based on efficient and high- speed electro -optic conversion. Nat Commun 16, 8178 (2025)

2025
[22]

Ou, S. et al. Hypermultiplexed integrated photonics –based optical tensor processor. Science Advances 11, eadu0228 (2025)

2025
[23]

Wang, J. et al. Microcomb-enabled parallel self- calibration optical convolution streaming processor. Light Sci Appl 15, 149 (2026)

2026
[24]

Wu, Y . et al. Integrated photonic 3D tensor processing engine. Light Sci Appl 15, 154 (2026)

2026
[25]

Meng, X. et al. High-integrated photonic tensor core utilizing high-dimensional lightwave and microwave multidomain multiplexing. Light Sci Appl 14, 27 (2025)

2025
[26]

Khaled, A., Aadhi, A., Huang, C., Tait, A. N. & Shastri, B. J. Fully integrated hybrid multimode -multiwavelength photonic processor with picosecond latency. Nat Commun 17, 28 (2025)

2025
[27]

Xing, S. et al. Seamless optical cloud computing across edge -metro network for generative AI. Nat Commun 16, 6097 (2025)

2025
[28]

N., Nahmias, M

Tait, A. N., Nahmias, M. A., Shastri, B. J. & Prucnal, P. R. Broadcast and Weight: An Integrated Network For Scalable Photonic Spike Processing. J. Lightwave Technol. 32, 4029–4041 (2014)

2014
[29]

Huang, C. et al. A silicon photonic–electronic neural network for fibre nonlinearity compensation. Nat Electron 4, 837– 844 (2021)

2021
[30]

Feldmann, J. et al. Parallel convolutional processing using an integrated photonic tensor core. Nature 589, 52–58 (2021). 17

2021
[31]

Zhou, W. et al. In-memory photonic dot-product engine with electrically programmable weight banks. Nat Commun 14, 2887 (2023)

2023
[32]

Xu, X. et al. 11 TOPS photonic convolutional accelerator for optical neural networks. Nature 589, 44–51 (2021)

2021
[33]

Chen, Z. et al. Deep learning with coherent VCSEL neural networks. Nat. Photon. 17, 723–730 (2023)

2023
[34]

Yin, R. et al. Integrated WDM -compatible optical mode division multiplexing neural network accelerator. Optica, OPTICA 10, 1709–1718 (2023)

2023
[35]

Sun, S. et al. Highly efficient photonic convolver via lossless mode-division fan-in. Nat Commun 16, 7513 (2025)

2025
[36]

Ling, Q. et al. On-chip optical matrix-vector multiplier based on mode division multiplexing. Chip 2, 100061 (2023)

2023
[37]

Yu, X. et al. Parallel optical computing capable of 100-wavelength multiplexing. eLight 5, 10 (2025)

2025
[38]

Yang, K. Y . et al. Multi-dimensional data transmission using inverse -designed silicon photonics and microcombs. Nat Commun 13, 7862 (2022)

2022
[39]

Nikkhah, V . et al. Inverse-designed low -index-contrast structures on a silicon photonics platform for vector –matrix multiplication. Nat. Photon. 18, 501–508 (2024)

2024
[40]

Sun, A. et al. Edge-guided inverse design of digital metamaterial- based mode multiplexers for high- capacity multi- dimensional optical interconnect. Nat Commun 16, 2372 (2025)

2025
[41]

Piggott, A. Y . et al. Inverse design and demonstration of a compact and broadband on- chip wavelength demultiplexer. Nature Photon 9, 374–377 (2015)

2015
[42]

Piggott, A. Y . et al. Inverse-Designed Photonics for Semiconductor Foundries. ACS Photonics 7, 569–575 (2020)

2020
[43]

Bian, Y . et al. 300-mm monolithic CMOS silicon photonics foundry technology [Invited]. in CLEO 2024 (2024), paper ATu3H.1 A Tu3H.1 (Optica Publishing Group, 2024). doi:10.1364/CLEO_AT.2024.A Tu3H.1

work page doi:10.1364/cleo_at.2024.a 2024
[44]

Banat, M. M. & Kavehrad, M. Reduction of optical beat interference in SCM/WDMA networks using pseudorandom phase modulation. J. Lightwave Technol. 12, 1863–1868 (1994)

1994
[45]

Dong, B. et al. Partial coherence enhances parallelized photonic computing. Nature 632, 55–62 (2024)

2024
[46]

F., Biehl, L

Baumgardner, M. F., Biehl, L. L. & Landgrebe, D. A. 220 Band A VIRIS Hyperspectral Image Data Set: June 12, 1992 Indian Pine Test Site 3. https://doi.org/10.4231/R7RX991C doi:10.4231/R7RX991C

work page doi:10.4231/r7rx991c 1992
[47]

Mourgias-Alexandris, G. et al. Noise-resilient and high -speed deep learning with coherent silicon photonics. Nat Commun 13, 5572 (2022)

2022
[48]

https://engineering.case.edu/bearingdatacenter

Bearing Data Center | Case School of Engineering. https://engineering.case.edu/bearingdatacenter
[49]

Han, X. et al. Breaking dense integration limits: inverse -designed lithium niobate multimode photonic circuits. Nat Commun 17, 1162 (2025)

2025
[50]

L., Dalvand, N

Pita Ruiz, J. L., Dalvand, N. & Ménard, M. Integrated silicon nitride devices via inverse design. Nat Commun 16, 9307 (2025)

2025
[51]

Shen, R. et al. 100 nm broadband and ultra-compact multi-dimensional multiplexed photonic integrated circuit for high- capacity optical interconnects. Photon. Res., PRJ 12, 2891–2900 (2024)

2024
[52]

Li, Q. et al. Ultra-broadband near- to mid-infrared electro-optic modulator on thin-film lithium niobate. Nat Commun 17, 1138 (2026)

2026
[53]

Rizzo, A. et al. Massively scalable Kerr comb-driven silicon photonic link. Nat. Photon. 17, 781–790 (2023)

2023
[54]

Song, X. et al. Full dimensional control of structured microwaves based on microcombs. Nat. Photon. 1–9 (2026) doi:10.1038/s41566-026-01843-0

work page doi:10.1038/s41566-026-01843-0 2026
[55]

Pirmoradi, A. et al. Integrated multi-port multi-wavelength coherent optical source for beyond Tb/s optical links. Nat Commun 16, 6387 (2025)

2025
[56]

Lischke, S. et al. Ultra-fast germanium photodiode with 3-dB bandwidth of 265 GHz. Nat. Photon. 15, 925–931 (2021)

2021
[57]

Steckler, D. et al. Monolithic electro-optic platform on silicon with bandwidth of 100 GHz and beyond. Nat Commun 16, 10789 (2025)

2025
[58]

Han, C. et al. Exploring 400 Gbps/λ and beyond with AI-accelerated silicon photonic slow-light technology. Nat Commun 16, 6547 (2025)

2025
[59]

Daudlin, S. et al. Three-dimensional photonic integration for ultra-low-energy, high-bandwidth interchip data links. Nat. 18 Photon. 19, 502–509 (2025)

2025
[60]

Yamazaki, H. et al. Analog Electronic and Optical Multiplexing Techniques for Transmitter Bandwidth Extension. Journal of Lightwave Technology 43, 1550–1564 (2025)

2025
[61]

Lu, H.- Y. et al. Low-loss High -uniformity Silicon Nitride Optical Building Blocks Integrated on Silicon Photonics Platform. in Optical Fiber Communication Conference (OFC) 2025 (2025), paper Th1G.2 Th1G.2 (Optica Publishing Group, 2025). doi:10.1364/OFC.2025.Th1G.2

work page doi:10.1364/ofc.2025.th1g.2 2025
[62]

Feng, H. et al. Integrated lithium niobate microwave photonic processing engine. Nature 627, 80–87 (2024)

2024
[63]

Deng, X. et al. Inverse Design of a Wavelength (De)Multiplexer for 1.55- and 2-μm Wavebands by Using a Hybrid Analog-Digital Method. Journal of Lightwave Technology 42, 5231–5240 (2024)

2024
[64]

Sun, A. et al. Inverse design of an ultra -compact dual-band wavelength demultiplexing power splitter with detailed analysis of hyperparameters. Opt. Express 31, 25415 (2023)

2023
[65]

Zhang, W. et al. Silicon microring synapses enable photonic deep learning beyond 9 -bit precision. Optica, OPTICA 9, 579–584 (2022)

2022