⚡ electrical engineering

Fully multiplexed photonic tensor computing

The paper introduces FieldCore, a fully multiplexed photonic tensor core that leverages inverse-designed silicon photonics to simultaneously harness five optical dimensions, achieving an estimated 69.12 TOPS throughput and enabling high-fidelity, parallel processing for diverse computational workloads.

Original authors: Junwen Zhang, Aolong Sun, Junhao Zhao, Fangchen Hu, Sizhe Xing, Yuqin Yuan, Jialin He, Yongzhu Hu, Xuyu Deng, Yinjun Liu, Ouhan Huang, Baiheng Zhao, Hancheng Liu, Tian Dong, Jingkai Zhou, Haoyang Sun

Published 2026-07-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Junwen Zhang, Aolong Sun, Junhao Zhao, Fangchen Hu, Sizhe Xing, Yuqin Yuan, Jialin He, Yongzhu Hu, Xuyu Deng, 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

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world's computers as giant, busy kitchens. Right now, these kitchens are getting crowded. They are trying to cook massive amounts of data for artificial intelligence, but the chefs (processors) are stuck waiting for ingredients to be moved from the pantry to the stove. This "traffic jam" of moving data is slowing everything down and using up a lot of energy. Scientists are looking for a new way to cook that doesn't rely on moving ingredients around so much. They are turning to light.

Think of light not just as a single beam, but as a super-flexible highway. In a normal computer, data travels like cars on a single-lane road, one after another. But in the world of photonics (using light to compute), scientists realized that light has many hidden lanes. You can send different messages on different colors of light at the same time (like different radio stations), on different "vibrations" of the light beam, or even by splitting the beam into different paths inside a tiny chip. The big question has been: Can we build a computer that uses all these lanes at once, perfectly in sync, to do complex math without the traffic jam?

This is exactly what a team of researchers from Fudan University and other institutions set out to do. They introduced a new invention called FieldCore, which they describe as a "fully multiplexed photonic tensor core." In simple terms, a "tensor" is just a fancy math word for a multi-dimensional block of data (like a stack of images or a 3D grid of numbers), and "multiplexed" means sending many things at once.

The researchers built a tiny silicon chip that acts like a master conductor for light. Instead of just using one color of light or one path, FieldCore grabs data and spreads it out across five different dimensions simultaneously: time (when the signal arrives), space (which path it takes), wavelength (the color of the light), radio-frequency (the speed of the signal's vibration), and guided-mode (the shape of the light wave).

Imagine a single stream of water hitting a magical nozzle. Instead of just spraying one jet, this nozzle instantly turns that water into a massive, organized grid of thousands of tiny, independent streams, all flowing through the same pipe. FieldCore does this with light. It takes a single input and expands it into a huge, parallel army of data streams. The team showed that this chip can perform math operations (specifically multiplying and adding numbers, which is the core of AI) on all these streams at the exact same time, without them getting mixed up or losing their shape.

They tested this "super-chip" with some impressive tasks. They made it recognize handwritten numbers (like the digit '9') with high accuracy, even when processing hundreds of images at once. They also used it to analyze complex images of the Earth's surface to identify different types of land and to listen to the vibrations of machines to spot faults before they break. The results were strong: the chip handled data at incredibly fast speeds (up to 120 GBaud) and maintained high precision, meaning the math was correct even with all that parallel traffic.

The team estimates that if they fully maximize this design, a single FieldCore chip could perform about 69.12 tera operations per second (TOPS) and handle up to 1,800 parallel input streams at once. This isn't just a small improvement; it's a completely new way of thinking about how to process information. By using the full "power" of light's many dimensions, FieldCore offers a promising path to building faster, more efficient AI computers that don't get stuck in the data traffic jams of today's electronics.

Technical Summary: Fully Multiplexed Photonic Tensor Computing (FieldCore)

Problem Statement

Tensor operations, particularly matrix multiplications and convolutions, are the computational backbone of modern artificial intelligence and high-throughput signal processing. While electronic accelerators have scaled through replicated arrays of processing units, they face increasing bottlenecks related to data movement costs across memory hierarchies and interconnects. Photonic computing offers a promising alternative by leveraging the inherent parallelism of optical fields. However, existing photonic processors typically exploit only isolated dimensions (e.g., wavelength-division multiplexing or time-domain processing) or partially combined dimensions. A critical challenge remains: as more optical dimensions are engaged, imperfections such as non-orthogonality, dispersion, and fabrication non-idealities break the uniformity of computation across channels. Consequently, there is a lack of a general framework that can systematically harness all native multiplexing dimensions of guided optical signals (wavelength, radio-frequency, guided-mode, space, and time) simultaneously while preserving high-fidelity, uniform computation across the entire signal space.

Methodology

The authors introduce FieldCore, a fully multiplexed photonic tensor core designed to overcome these limitations by jointly harnessing five dimensions: wavelength, radio-frequency (RF) subcarrier, guided-mode, space, and time.

Architecture and Principle

FieldCore operates on the principle that a guided optical field can be expanded across multiple orthogonal dimensions to form a high-order tensor. The system maps input data onto these dimensions, allowing a single photonic chip to perform multiplicative parallelism.

Signal Representation: The optical field is represented as a sum over optical carrier frequencies ( $\omega_\ell$ ), vectorial guided eigenmodes ( $\psi_m$ ), and RF subcarrier offsets ( $\Omega_n$ ).
Hardware Implementation: The core is implemented as a $K \times R$ crossbar on a silicon-on-insulator (SOI) platform. It integrates inverse-designed building blocks to ensure compactness, broadband response, and fabrication tolerance.
Key Components:
- Inverse-Designed Digital-Metamaterials: Used for mode multiplexers/demultiplexers (MUX/deMUX), 3-dB splitters, and waveguide crossings. These enable low-crosstalk, broadband operation in a $\mu$ m-scale footprint.
- Mode-Insensitive Phase Shifters (MIPS): Engineered by widening waveguides to equalize the thermo-optic response of TE0 and TE1 modes, ensuring weight invariance across the mode dimension.
- Programmable Routing and Weighting: Tunable couplers (dual-port Mach-Zehnder interferometers) and multipliers (single-port MZI attenuators) distribute signals and apply analog weights.
- Accumulation: Tunable adders perform weighted summation across spatial input channels.

Operational Flow

Input: Temporal streams are up-converted to RF subcarriers, modulated onto specific wavelengths, and mapped to guided modes via on-chip MUX.
Processing: Signals are routed through the crossbar where they undergo parallel weighted accumulation. The same weight matrix is applied uniformly across all multiplexed slices (RF, wavelength, mode, time).
Output: Multiplexing is reversed in the optical and electrical domains to recover parallel tensor slices.

Key Contributions

Unified Multi-Dimensional Tensor Core: FieldCore is the first integrated photonic tensor core to unify wavelength, RF subcarrier, guided mode, space, and time within a single framework, converting native parallel dimensions into multiplicative tensor-level parallelism.
Dimension-Uniform Computation: The design achieves high-fidelity computation across all dimensions. This is validated by:
- Mode Invariance: Using MIPS to maintain consistent weighting for TE0 and TE1 modes.
- Broadband Flatness: Ensuring spectrally flat responses for wavelength and RF dimensions.
Inverse-Designed Photonic Building Blocks: The use of inverse-designed digital-metamaterial components allows for compact, broadband (>40 nm), and low-crosstalk (< -20 dB) multimode operation, essential for maintaining channel consistency.
Scalable Architecture: The system supports up to 1,800 parallel input streams within a single core, establishing a scalable paradigm for photonic tensor computing.

Experimental Results

The authors fabricated a $4 \times 4$ FieldCore chip (2.55 $\times$ 2.58 mm $^2$ ) and validated its performance through arithmetic verification and complex inference tasks.

Arithmetic Verification

Speed and Precision: The system demonstrated multiply-accumulate (MAC) operations with over 5-bit precision at symbol rates up to 120 GBaud for both TE0 and TE1 modes.
Wavelength Stability: Precision variations across the C-band (1530–1565 nm) were below 0.7 bits relative to the 1550 nm reference.
RF Scalability: The system supported up to 100 RF subcarriers on a single optical carrier with an aggregate baudrate of 50 GBaud, maintaining ~5-bit addition precision.

Task-Level Demonstrations

Image Convolution:
- Grayscale: Performed directional edge extraction on a 150 $\times$ 150 image using four 2 $\times$ 2 kernels, achieving a peak signal-to-noise ratio (PSNR) of 30.5 dB at 10 GBaud and 23.4 dB at 70 GBaud.
- RGB: Successfully transmitted and convolved RGB images using FDM, preserving structural details across channels and modes.
MNIST Handwritten Digit Recognition:
- Utilized joint wavelength, mode, and RF multiplexing (200 parallel data lanes) to process handwritten digits.
- Achieved an average accuracy of 93.8% across all multiplexed channels, compared to a 95.0% digital baseline.

High-Dimensional and Parallel Applications

Channel-Parallel Hyperspectral Classification: Mapped 200 spectral bands of the Indian Pines dataset to 200 parallel channels. The system achieved an overall classification accuracy of 90.8% across 12 land-cover classes, corresponding to an equivalent analog precision of 5.16 bits.
Instance-Parallel Mechanical Fault Diagnosis: Processed 200 concurrent vibration signal instances from the CWRU bearing dataset. The system achieved an overall accuracy of 95.1% across seven operating conditions (normal and six fault types), with robust per-channel performance.

Significance and Claims

The paper claims that FieldCore establishes a scalable route towards high-throughput and massively parallel photonic tensor processing on chip. By utilizing all native multiplexing dimensions of guided optical signals, FieldCore addresses the limitations of current electronic and partially multiplexed photonic systems.

Throughput: The authors estimate an aggregate compute throughput of 69.12 Tera Operations Per Second (TOPS) in a throughput-maximized regime.
Parallelism: The architecture can accommodate up to 1,800 parallel input streams within a single core, a level of native parallelism difficult to achieve in conventional electronic processors which rely on core replication or time-sequenced scheduling.
Generalizability: The work provides a general framework for fully multiplexed photonic computing that can be extended to other architectures and material systems, offering a path to overcome the data movement bottlenecks of modern AI workloads.

The authors emphasize that the system preserves a uniform programmed computation across all multiplexed channels, a critical requirement for high-fidelity tensor processing that previous multi-dimensional demonstrations struggled to maintain.