Papers
Topics
Authors
Recent
Search
2000 character limit reached

Fully multiplexed photonic tensor computing

Published 24 Apr 2026 in physics.optics | (2604.22660v1)

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.

Summary

  • The paper introduces FieldCore, a photonic tensor core that fully multiplexes across wavelength, RF, mode, time, and space for scalable tensor operations.
  • It demonstrates high-fidelity analog processing with <1% weight mismatch, over 40 nm bandwidth, and precision exceeding 5 bits at symbol rates up to 120 GBaud.
  • Experimental results validate robust image convolution and AI inference, achieving 93.8% digit recognition accuracy and 69.12 TOPS throughput.

Fully Multiplexed Photonic Tensor Computing: Architecture, Implementation, and Application

Motivation and Context

The acceleration of tensor operations is a fundamental challenge in AI, scientific computing, and high-throughput signal processing. Traditional electronic tensor accelerators, while having achieved considerable parallelism through replicated cores and advanced memory hierarchies, are increasingly limited by inefficiencies in data movement and constrained computational density. Photonic computing offers intrinsic parallelism with guided optical fields, allowing encoding and processing across mutually orthogonal channels—wavelength, guided modes, radio-frequency (RF) subcarriers, time, and space. However, prior photonic processors have typically exploited only a subset of these dimensions, and stacking multiplexing techniques has proven inadequate due to breaking of operator uniformity by imperfection and non-idealities.

FieldCore Architecture and Design

FieldCore is presented as a fully multiplexed photonic tensor core leveraging the complete multidimensional space of guided optical fields. The core processes tensor data in parallel across wavelength, RF subcarrier, guided mode, time, and space, enabling multiplicative scaling of parallelism. The hardware instantiation is a compact silicon-photonic chip integrating inverse-designed building blocks—mode multiplexers/demultiplexers (MUX/deMUX), tunable couplers, low-crosstalk crossings, analog multipliers, and tunable adders.

Inverse-designed digital metamaterial structures underpin the broadband, low-crosstalk, and footprint-efficient operation. Mode invariance is engineered using thermo-optically matched phase shifters, while wavelength and RF invariance are ensured through a spectrally flat response. The FieldCore as a K×RK \times R crossbar applies uniform weight matrices across all multiplexed streams, supporting a full tensor operation spanning dimensions indexed by RF subcarrier count (NfN_f), wavelength (NaN_a), guided mode (NmN_m), and time (TT).

Experimental Validation: Multi-Dimensional Arithmetic Operations

Device-level characterization demonstrates high-fidelity analog weight loading (<1% mismatch), broadband operation (>40 nm), low inter-mode crosstalk (< -20 dB), and mode-consistent weighting. The multiply-accumulate (MAC) operation is experimentally validated across mode (TE0, TE1), wavelength, and RF dimensions with precision exceeding 5 bits at symbol rates up to 120 GBaud. Precision variation across the C band is below 0.7 bits, and RF parallelization supports up to 100 subcarriers per optical carrier—showing that parallelism in the FieldCore is currently analog bandwidth-limited, not subcarrier-limited.

Application Demonstrations: Image Convolution and AI Inference

FieldCore supports high-speed convolution in mode and RF dimensions with high-fidelity outcomes:

  • Grayscale and RGB image convolution: PSNR values of 30.11 dB and 28.54 dB, respectively, confirm analog weight programmability and multiplexed fidelity.
  • Parallel MNIST digit recognition: Two wavelength groups, two modes, and 50 RF subcarriers enable 200 parallel input streams in a single shot. Experimental accuracy of 93.8% matches the digital baseline (95.0%), confirming robustness under fully multiplexed computation.

High-Dimensional and Massively Parallel AI Inference

FieldCore unlocks two application regimes within a single core:

  • Channel-Parallel Hyperspectral Image Classification: The Indian Pines dataset with 200 spectral bands is mapped one-to-one onto 200 FieldCore channels. Four shared kernels perform spatial convolution in parallel across all channels. Experimental overall classification accuracy reaches 90.8% (12 classes), with precision sweeps demonstrating robustness down to 4-bit analog precision.
  • Instance-Parallel Mechanical Fault Diagnosis: Parallel processing of 200 signal instances from the CWRU bearing dataset achieves 95.1% overall accuracy across seven fault classes, with per-channel accuracy above 90%. t-SNE analysis shows inter-class separation, demonstrating the quality of extracted features.

FieldCore supports an estimated compute throughput of 69.12 TOPS, with up to 1,800 parallel input streams in throughput-optimized regimes—several orders of magnitude beyond typical electronic tensor accelerators.

Implications and Future Directions

The architecture advances integrated photonic tensor cores by unifying wavelength, RF, mode, space, and time dimensions for on-chip computation, surpassing previous approaches that were limited to combined or isolated multiplexing (2604.22660). Notably, FieldCore establishes channel-uniformity and fidelity under dense parallelization—addressing critical barriers in photonic computing.

Future scaling opportunities are outlined:

  • Modal space expansion with inverse-designed multimode/polarization devices.
  • Wavelength scaling via broadband weight units and integrated optical combs.
  • RF scaling through co-integrated high-bandwidth photonics and electronic front-ends.
  • Foundry improvements for lower waveguide losses and higher uniformity, enabling larger core arrays.

The fully multiplexed photonic computing strategy provides a generalized route for scalable AI inference and scientific workloads, and is extendable to various photonic computing platforms and material systems.

Conclusion

FieldCore demonstrates fully multiplexed photonic tensor computing leveraging all native multiplexing dimensions of the optical field. Robust experimental performance spanning arithmetic precision, convolution, and large-scale parallel AI inference is attained with high fidelity, high throughput, and multiplicative scaling of parallelism. This architectural approach establishes a practical and theoretically scalable paradigm for photonic tensor accelerators, with direct implications for high-dimensional, real-time AI inference and scientific data processing. Future developments will likely focus on further modal, wavelength, and RF scaling, as well as system-level integration with emerging photonic foundry technologies.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Sign Up to Generate All Videos Subscribe on YouTube

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Sign Up to Generate

Explain it Like I'm 14

What is this paper about?

This paper introduces FieldCore, a tiny silicon chip that does math with light instead of electricity. Its big idea is to use many “lanes” inside a single beam of light—different colors, radio-like subchannels, shapes of light in a waveguide, time, and separate input paths—all at once. By doing that, the chip can process lots of data in parallel, very fast, using the same operation on every lane.

What questions were the researchers trying to answer?

The team set out to answer:

  • Can we build one photonic “tensor core” that uses all of light’s natural ways to carry information (color, RF subcarriers, modes, time, and space) at the same time?
  • Can this core run the exact same programmed calculation across all those lanes with high accuracy?
  • Can it handle real tasks, like image filtering and recognizing digits, not just simple lab demos?
  • How fast and how parallel can such a chip be—compared to today’s electronics?

How did they build and test it?

Think of a busy multi-lane freeway:

  • Different colors of light = cars in different colored lanes (wavelengths).
  • Radio-frequency subcarriers = cars in the same lane but listening to different radio stations.
  • Guided modes = different “layers” or shapes of traffic stacked in the same road.
  • Time = cars passing through over time.
  • Space = separate on-chip entry points.

FieldCore is like a toll plaza that applies the same “toll rule” (a weighted sum) to every car in every lane, all at once.

How it works (in simple terms):

  • The chip is a crossbar: many inputs go to many outputs through adjustable “weights.” This is the basic building block for matrix multiplication and convolution in AI.
  • Data are loaded onto many lanes of light: different colors, different RF subcarriers on each color, and two “modes” (two shapes of light) inside the same waveguide. Multiple time steps stream through.
  • Special tiny optical components—designed by computer (“inverse design”)—split, combine, and steer light with very low interference and wide bandwidth.
  • Adjustable optical elements (mini interferometers and heaters) set the numeric weights for the math. The design makes sure the weights act the same on both light modes (so every lane gets the same calculation).
  • After the math happens optically, the results are read back out by separating the lanes and converting the light back into electrical signals.

They verified three things:

  1. The basic math (multiply, add, and multiply–accumulate) works accurately at high speed.
  2. The operations stay the same across modes, colors, and RF subcarriers (so every lane is treated equally).
  3. The system solves real tasks: image edge detection, RGB image processing, handwritten digit recognition, hyperspectral image classification, and machine fault detection.

What did they find, and why is it important?

Main results:

  • Same operation across all lanes: The chip kept the same programmed weight on different modes (two light “shapes”), across many wavelengths (colors), and across up to 100 RF subcarriers on one color. That’s crucial for doing identical AI computations in true parallel.
  • High speed and solid accuracy: They ran multiply–accumulate up to 120 GBaud (very fast symbol rate) with over 5-bit equivalent precision. In practice, that’s good enough for many AI inference tasks that tolerate limited precision.
  • Real tasks worked well:
    • Image convolution (edge detection) on grayscale and RGB images, keeping good image quality even at high speeds.
    • MNIST handwritten-digit recognition, using 200 parallel lanes at once, reached 93.8% accuracy (close to a 95.0% digital baseline).
    • Hyperspectral land-cover classification (200 spectral bands) hit 90.8% accuracy.
    • Mechanical fault diagnosis from vibration signals (7 classes) hit 95.1% accuracy, with cleanly separated feature clusters.
  • Big parallelism and throughput: In estimates based on the experiments, FieldCore can handle up to 1,800 parallel input streams in one core and reach about 69.12 tera operations per second (TOPS). That’s a lot of simultaneous processing for one tiny chip.

Why it matters:

  • Today’s AI chips are hitting limits because moving data around and adding more cores is getting harder and more power-hungry. Light can carry many data streams at once in a single waveguide, naturally boosting parallelism without just copying more hardware.
  • FieldCore shows you can program one computation and run it across many optical lanes uniformly, which is essential for reliable, scalable AI inference.
  • This approach could reduce data-movement bottlenecks and enable fast, parallel processing in places like data centers, satellites (for hyperspectral imaging), and factory edge nodes (for machine health monitoring).

What could this lead to?

  • Scaling up: More modes (including different polarizations), more colors (wider optical bands), and more RF subcarriers could further enlarge parallelism as photonics and electronics improve.
  • Practical systems: With better integrated lasers, modulators, detectors, and co-packaged electronics, chips like FieldCore could become compact accelerators for AI tasks that need to process many data streams at once.
  • New applications: Anywhere there’s a flood of high-dimensional data—like Earth observation, industrial IoT, or large sensor networks—fully multiplexed photonic cores could perform fast, parallel computations at low latency.

Key takeaways

  • FieldCore is a photonic “tensor core” that uses light’s multiple lanes—color, RF subcarriers, guided modes, time, and space—to do the same math in massive parallel.
  • It keeps operations uniform across all lanes, runs at very high symbol rates with ~5-bit precision, and succeeds on real AI tasks.
  • It demonstrates a path toward scalable, high-throughput, parallel AI inference that’s hard to achieve with conventional electronics alone.
View Paper Prompt  View All Prompts 

Knowledge Gaps

Knowledge gaps, limitations, and open questions

Below is a focused list of what remains missing, uncertain, or unexplored, framed to be actionable for follow-on research:

  • End-to-end parallel I/O not demonstrated: on-chip computation is simultaneous, but channel-resolved readout used a shared receiver and sequential acquisition; a fully parallel demultiplexing and detection/readout stack (per-wavelength, per-mode, per-subcarrier) remains to be built and evaluated.
  • System-level energy and TOPS/W unknown: no measurement of energy per MAC including modulators, drivers, heaters, EDFAs, photodetectors, and DSP; provide full-stack power/throughput benchmarks and compare against electronic accelerators.
  • Weight storage and reconfiguration are thermal and slow: TiN heater-based MZIs imply ms-scale updates, static power, and drift; quantify re-tuning time, retention, and drift over hours/days and explore low-power, fast, or non-volatile weights (carrier-depletion, barium titanate, phase-change materials).
  • Uniform-broadcast weight constraint: the same kernel is broadcast across all multiplexed channels; architectures for per-channel (per-wavelength/mode/subcarrier) independent weight programming without sacrificing multiplexing are needed for more general workloads.
  • Limited mode count: only TE0/TE1 used; scaling to many spatial modes and dual polarization requires quantifying mode-dependent loss, crosstalk, and group-delay mismatch, and co-designing mode-MUX/deMUX and mode-insensitive phase shifters for >2 modes and both polarizations.
  • WDM scalability not established: only eight wavelengths were used; characterize inter-wavelength beating under direct detection as channel count and RF bandwidth grow, OSNR/ASE penalties, filter requirements, and feasibility of comb sources; compare with coherent detection and balanced PD to mitigate beating.
  • RF subcarrier scaling and linearity: while up to 100 subcarriers were tested at reduced per-lane baud, intermodulation distortion, PAPR/crest-factor effects, mixer/LO phase noise, and guard-band design for higher aggregate bandwidths remain uncharacterized.
  • Dispersion and group delay: quantify chromatic dispersion and inter-mode group delay across C-band and two modes at 50–120 GBaud, and determine equalization needs for symbol alignment across multiplexed dimensions.
  • Dynamic range and linearity of analog compute: measure SFDR/IM3 of MZI multipliers and adders, receiver linearity, and distortion of FDM summation under realistic multi-tone loading and temperature variation.
  • Signed-weight implementation overhead: the reference-offset (differential) mapping doubles resources and tightens dynamic-range/SNR; quantify accuracy/SNR penalties and explore architectures enabling native signed (or complex) weights.
  • Scaling of crossbar size: only a 4×4 core was realized; analyze and experimentally validate loss, SNR, crosstalk, and accuracy vs K×R for larger arrays (e.g., ≥16×16), including power-splitting dilution and accumulation-path loss.
  • Loss and noise budget: the experiments rely on EDFAs and ~5 dB/facet grating coupling; provide a detailed OSNR budget per channel as channel count grows and assess alternatives (edge couplers, on-chip lasers/amplifiers) to remove EDFAs and reduce ASE noise.
  • Long-term stability and calibration: no closed-loop calibration shown; develop in-situ monitors and control (dithering/locking) to track thermal drift, device aging, and process variation, and report calibration overhead at scale.
  • Real-time operation and DSP cost: most tasks used offline DSP/oscilloscope capture; demonstrate real-time FDM demodulation, synchronization, and classification on hardware and quantify the DSP and ADC energy/latency per subcarrier.
  • Precision limits for deep models: ~5-bit equivalent precision was validated on small CNNs; systematically benchmark larger/deeper networks (e.g., ResNet/MobileNet/Transformers) with quantization-/noise-aware training to identify minimum precision and accuracy scaling.
  • On-chip nonlinearity and multi-layer optical stacking: activations were electronic; assess the latency/energy of repeated O/E–E/O conversions between layers and explore integrated optical/nonlinear devices to reduce conversions or enable optical activations.
  • Throughput projections unvalidated at scale: 69.12 TOPS and 1,800 input streams are projections; perform a full demonstration approaching these limits with simultaneous multi-dimension I/O and report bottlenecks.
  • Per-dimension variability characterization: provide full distributions (not only averages) of precision/PSNR vs wavelength, mode, and subcarrier to identify worst-case channels and devise equalization or channel-assignment strategies.
  • Mapping generality beyond 2×2 kernels: extend demonstrations to larger kernels, strides/dilations, depthwise/pointwise convs, and GEMM; quantify weight reuse, routing overhead, and scheduling for realistic CNN layers.
  • Data movement analysis: quantify costs of electrical-to-optical mapping (DACs, modulators) and optical-to-electrical recovery (PDs, ADCs) relative to compute savings, especially when scaling RF/WDM channel counts.
  • Heater power and thermal crosstalk at scale: measure per-weight static/dynamic power, thermal time constants, and inter-heater coupling across dense arrays; assess feasibility for thousands of weights.
  • Fabrication yield and variability: inverse-designed components showed robustness in simulation, but wafer-scale statistics across many dies and long-term reliability (heaters, metals) are missing.
  • Polarization handling and packaging: only TE grating couplers were used; demonstrate polarization diversity, edge-coupled packaging with lower loss, and tolerance to polarization rotation in deployed fiber links.
  • Coherent vs direct detection trade-offs: open question whether moving to coherent detection could raise precision, enable signed/complex weights, and ease WDM scaling versus added complexity and power.
  • Integration with active photonics and CMOS: current setup uses off-chip MZMs/PDs/AWGs; demonstrate monolithic or co-packaged modulators, detectors, drivers, and multicarrier RFICs, and evaluate co-integration limits (bandwidth, crosstalk, heat).
  • Security/robustness to environmental perturbations: assess sensitivity to laser frequency drift, vibration, and temperature fluctuations across multiplexed dimensions; design mitigation (locking, redundancy).
  • Benchmarking against electronic baselines: provide fair, task-level throughput/latency/energy comparisons including I/O and memory traffic for representative workloads beyond MNIST and small patches (e.g., ImageNet-scale inference, streaming analytics).
View Paper Prompt  View All Prompts 

Practical Applications

Overview

FieldCore is a fully multiplexed photonic tensor core that executes the same analog weighted-sum operation across wavelength, RF subcarriers, guided modes, space, and time within a single optical field. The chip (silicon photonics, inverse-designed, MZI-based weighting and summation) demonstrated >5-bit precision MACs up to 120 GBaud, parallel image convolutions, MNIST inference (93.8% across 200 channels), hyperspectral classification (90.8% accuracy across 200 bands), and mechanical fault diagnosis (95.1% across 200 simultaneous instances). It supports an estimated 69.12 TOPS and up to ~1,800 parallel input streams.

Below are actionable applications tied to real-world sectors, each labeled by timeframe and accompanied by likely tools/products/workflows and feasibility assumptions.

Immediate Applications

These can be piloted or deployed in controlled environments using today’s silicon-photonics packaging, external lasers/modulators, and hybrid electronic readout, leveraging 4–5 bit-equivalent precision and hundreds of concurrent channels.

  • Industrial predictive maintenance edge nodes for parallel vibration analytics (Industry: manufacturing, energy, logistics)
    • Use case: Concurrently analyze 100–200+ machine vibration streams for fault detection (as shown with CWRU dataset, 95.1% accuracy) to reduce downtime and maintenance cost.
    • Tools/products/workflows: FieldCore-based photonic inference module; noise-aware training to 4–5 bits; pipeline to serialize 1×1024 waveforms to RF subcarriers; embedded post-processing classifier.
    • Assumptions/dependencies: Stable thermal control; calibration/drift management; integration with sensors and ADCs/DACs; acceptable model performance at 4–5 bits; controlled environmental conditions.
  • Hyperspectral analytics acceleration for remote sensing and lab instruments (Industry: Earth observation, agriculture, mining; Academia: remote sensing)
    • Use case: Channel-parallel convolution across 100–200 spectral bands to accelerate land-cover classification and materials mapping (as shown with Indian Pines, 90.8%).
    • Tools/products/workflows: “SpectralCore” lab accelerator; wavelength/RF mapping toolchain; noise-aware/quantization-aware training; APIs into GIS/remote sensing pipelines.
    • Assumptions/dependencies: External multi-wavelength sources (or combs) and filters; data ingest/egress bandwidth; robust calibration to maintain per-channel uniformity.
  • High-speed optical DSP kernels for coherent links and testbeds (Industry: telecom)
    • Use case: Photonic convolution/equalization (FIR-like filters) to offload linear DSP kernels (e.g., chromatic dispersion compensation, MIMO pre/post-processing) at high baudrates.
    • Tools/products/workflows: Plug-in photonic DSP module; weight-compilation from channel-estimate to MZI settings; control firmware for adaptive updates.
    • Assumptions/dependencies: Coherent Rx/Tx integration; precise timing and phase management; low-added latency; component bandwidth alignment; lab/pilot deployment initially.
  • Batch visual inspection pre-processing on production lines (Industry: semiconductor, pharma, electronics)
    • Use case: Real-time edge detection/convolution over many inspection streams in parallel to flag defects upstream of heavier compute.
    • Tools/products/workflows: Inline photonic preprocessor; 2×2/3×3 kernels mapped once across all channels; downstream GPU/CPU classifier.
    • Assumptions/dependencies: Low-loss coupling to cameras; tolerance to 4–5 bit convolution; environmental stability; throughput-balanced E/O and O/E.
  • Multichannel RF signal pre-processing for spectrum monitoring (Industry: wireless, defense, public safety)
    • Use case: Parallel channelization, filtering, and simple feature extraction across dozens-to-hundreds of RF bands using RF subcarriers and WDM.
    • Tools/products/workflows: RF-to-optical front-end; photonic filterbank; digital demux and lightweight classifier.
    • Assumptions/dependencies: High analog bandwidth E/O and O/E; calibration for amplitude/phase flatness; manageable crosstalk; compliance with RF front-end specs.
  • Academic photonic-AI testbed for algorithm–hardware co-design (Academia)
    • Use case: Research in low-precision analog photonic inference, quantization/noise-aware training, compiler mapping across wavelength/mode/RF/time, and inverse design validation.
    • Tools/products/workflows: Open-source FieldCore SDK (compiler, calibration, drift compensation, telemetry); datasets and repeatable benchmarks (MNIST, hyperspectral, vibration).
    • Assumptions/dependencies: Access to packaged chips and standard lab optics; institutional support for thermal control and calibration routines.
  • Photonic computing education and workforce training (Academia/Education)
    • Use case: Hands-on curricula covering multiplexed optical computing, MZI calibration, and AI workloads with analog photonic hardware.
    • Tools/products/workflows: Safe lab kits with low-power lasers; virtual/remote labs; courseware around FieldCore-style cores.
    • Assumptions/dependencies: Institutional lab safety; simplified tooling; reduced-cost demo hardware.
  • Early datacenter pilot for CNN front-end offload (Industry: cloud/software)
    • Use case: Offload first convolutional layers for large-batch inference where bandwidth-parallelism is exploitable and latency budgets are tight.
    • Tools/products/workflows: PCIe/CMC carrier with photonic core + control ASIC/FPGA; driver + runtime; noise-aware retraining; model partitioning.
    • Assumptions/dependencies: Packaging maturity; thermal/photonics control stack; quantization-tolerant models; ROI vs. GPUs in limited-scope pilots.
  • Standards and benchmarking initiatives for analog photonic inference (Policy/Industry consortia)
    • Use case: Establish precision metrics, datasets, and calibration protocols for analog photonic AI accelerators to enable fair comparison and procurement.
    • Tools/products/workflows: Community-driven benchmarks (e.g., 4–6 bit regimes, TOPS/W), channel-uniformity tests, drift/aging tests, conformance tests.
    • Assumptions/dependencies: Multi-stakeholder participation (vendors, labs, hyperscalers); open reporting of measurement methods.

Long-Term Applications

These require further integration (on-chip lasers/detectors/modulators), non-volatile/electro-optic weights, tighter thermal/phase control, broader WDM/RF/mode scaling, and mature toolchains.

  • Datacenter-scale photonic AI accelerators with co-packaged optics (Industry: cloud/AI, software)
    • Use case: High-throughput, low-latency inference for CNNs/transformers via massively parallel photonic tensor cores close to optical interconnects.
    • Tools/products/workflows: Co-packaged lasers/PDs; non-thermal phase shifters (carrier-depletion, EO materials); compiler/runtime with autotuning; photonic interconnects to memory/storage.
    • Assumptions/dependencies: Integration yield and reliability; >8-bit effective precision or architecture-level compensation; standardized APIs and orchestration; cost per TOPS advantage.
  • Onboard satellite and UAV hyperspectral payload processing (Industry: space, agriculture, disaster response; Policy: environmental monitoring)
    • Use case: Real-time spectral-feature extraction/classification at the sensor to reduce downlink and enable on-orbit decisions (e.g., crop stress, wildfire, mineral mapping).
    • Tools/products/workflows: Radiation-hardened photonic cores; temperature-compensated packaging; integrated frequency combs; model compression/quantization workflows.
    • Assumptions/dependencies: Space qualification (radiation, vibration, temperature); ultra-low-SWaP designs; autonomous calibration; long-term drift stability.
  • Edge AI coprocessors in robotics and autonomous vehicles (Industry: robotics, automotive)
    • Use case: Front-end convolution/feature extraction for cameras/LiDAR/event sensors with microsecond latency and high data rates.
    • Tools/products/workflows: Integrated PIC modules co-located with sensors; hybrid E/P pipelines; safety-certified software stacks.
    • Assumptions/dependencies: Ruggedized PICs; low power non-volatile weighting; reliable operation across temperature/EMI; standards for functional safety.
  • Healthcare imaging and signal-processing accelerators (Sector: healthcare/medtech)
    • Use case: Real-time OCT/OCTA pre-processing, ultrasound beamforming, ECG/EEG multichannel filtering and feature extraction with high channel counts.
    • Tools/products/workflows: Medical-grade PICs; secure/traceable calibration; clinical models trained for 4–6 bit hardware; integration with hospital IT.
    • Assumptions/dependencies: Regulatory approval; rigorous validation for diagnostic use; patient data privacy; long-term stability and sterilization constraints.
  • 6G RAN and massive-MIMO photonic baseband/equalization (Industry: telecom/wireless)
    • Use case: Photonic MIMO equalizers and filterbanks for wideband, many-antenna systems; joint RF/wavelength/mode parallelism for front-haul/back-haul.
    • Tools/products/workflows: Co-integration with RFICs and AFE; photonic-Ethernet fronthaul; dynamic weight updates from channel state.
    • Assumptions/dependencies: Tight synchronization; linearity and EVM targets; scalable calibration; competition with advanced CMOS DSP.
  • Consumer smart cameras and AR/VR wearables (Daily life/Consumer electronics)
    • Use case: Always-on photonic edge detection/denoising and low-latency feature extraction to reduce load on mobile SoCs and extend battery life.
    • Tools/products/workflows: PICs integrated near image sensors; ultra-low-power non-volatile weights; mobile SDKs for hybrid pipelines.
    • Assumptions/dependencies: Cost, footprint, and power advantages vs. CMOS; thermal management in compact enclosures; robust mass manufacturing.
  • Low-latency quantitative trading and real-time decision systems (Finance)
    • Use case: Photonic pre-processing and inference for ultra-low-latency signal detection/feature extraction across many feeds.
    • Tools/products/workflows: Datacenter co-location modules; deterministic latency pipelines; compliance logging and failover.
    • Assumptions/dependencies: Reliability and determinism under market stress; integration with existing FPGA stacks; regulatory compliance.
  • Power grid and industrial sensor network monitoring (Energy/Utilities)
    • Use case: Parallel time-series filtering and anomaly detection across thousands of sensors at substations or plants.
    • Tools/products/workflows: Rugged photonic inference appliances; digital twins and retraining pipelines; remote calibration/monitoring.
    • Assumptions/dependencies: Environmental hardening; cybersecurity; interoperability with SCADA; maintenance skillsets.
  • RF spectrum intelligence and electronic warfare support (Defense/Public safety)
    • Use case: Photonic channelization/classification of dense RF environments for real-time situational awareness and signal fingerprinting.
    • Tools/products/workflows: Wideband E/O front-ends; adaptive photonic filters; secure control and logging.
    • Assumptions/dependencies: Classified performance requirements; environmental resilience; long-term stability.
  • In-network photonic compute and compute-in-fiber (Industry: networking)
    • Use case: Offload linear transforms (e.g., filtering, projections) into the optical fabric within switches/links to reduce data movement.
    • Tools/products/workflows: Reconfigurable photonic cores within switches; orchestration APIs; telemetry for in-situ health.
    • Assumptions/dependencies: Network standards and management; fault isolation; alignment with co-packaged optics roadmaps.

Cross-cutting assumptions and dependencies

  • Photonic integration: On-chip or co-packaged lasers (e.g., combs), modulators, and detectors; low-loss packaging; polarization and mode management; scalable, foundry-compatible inverse-designed components.
  • Weight programmability: Migration from thermal phase shifters to low-power/high-speed mechanisms (carrier-depletion, barium titanate, lithium niobate, phase-change, MEMS) for efficiency and stability; non-volatile/fast update options.
  • Calibration and control: Robust, automated calibration, drift compensation, and health monitoring; temperature stabilization; per-channel uniformity across wavelength/RF/modes.
  • Software stack: Compilers to map tensors to multiplexing dimensions, noise-/quantization-aware training (4–6 bits), model partitioning, and runtime schedulers; standard APIs and benchmarks.
  • I/O and co-design: Efficient E/O and O/E interfaces to avoid data-movement bottlenecks; co-design with electronics for memory/IO; clear energy/TOPS and latency advantages vs. CMOS.
  • Application constraints: Tolerance for analog noise and limited precision; domain-specific validation (e.g., clinical trials, safety certifications, finance compliance); environmental ruggedization where required.
View Paper Prompt  View All Prompts 

Glossary

  • 3-dB splitter: A passive optical device that evenly splits optical power into two paths. "The mode multiplexer and the 3-dB splitter used in the optical multiplier were designed using analog-to-digital optimization methods"
  • amplified spontaneous emission (ASE) source: A broadband optical light source based on spontaneous emission that is amplified, used for spectral measurements. "launching broadband light from an amplified spontaneous emission source (OVLINK ASE- CL-PM)"
  • analog-to-digital optimization methods: Inverse-design approaches that start with continuous (analog) structures and map them to discrete (digital) patterns under fabrication constraints. "designed using analog-to-digital optimization methods39,62, which provide large topology freedom while naturally incorporating fabrication constraints"
  • arbitrary waveform generator (AWG): A programmable instrument that produces high-speed electrical waveforms for modulation. "driven directly by a 224 GSa s-1 arbitrary waveform generator (AWG, Keysight M8199B)"
  • B-spline parameterization: A smooth curve representation used here to optimize waveguide trajectories with low loss and crosstalk. "optimized as a continuous curve using a quasi-uniform B-spline parameterization"
  • crossbar: A matrix-like interconnect in which multiple inputs can be routed to multiple outputs for weighted accumulation. "At the hardware level, the FieldCore is implemented as a KxR crossbar"
  • deep trenches: Micromachined isolating gaps used to reduce thermal crosstalk between photonic elements. "100-um-long interferometer arms isolated by deep trenches for thermal-crosstalk suppression"
  • digital-metamaterial components: Photonic structures composed of discretized subwavelength pixels designed to achieve targeted optical functions. "inverse-designed digital-metamaterial components39, including the mode MUX/deMUX, the 3-dB splitter and the waveguide crossing"
  • direct detection: Optical reception technique that measures optical power without coherent mixing, simplifying receivers. "suppressing coherent interference during optical accumulation and direct detection"
  • electron-beam lithography (EBL): A high-resolution patterning technique for defining nanoscale photonic features. "two aligned electron-beam lithography (EBL) and inductively coupled plasma (ICP) etching steps"
  • erbium-doped fiber amplifier (EDFA): An optical amplifier that boosts signals in the C-band using erbium-doped fiber. "an erbium-doped fiber amplifier (EDFA, Amonics AEDFA-23-B-FA) was used to compensate insertion loss"
  • extinction ratio: The ratio between maximum and minimum transmitted power, indicating modulation or attenuation contrast. "The measured extinction ratios reach 30.90 dB for TE0 and 29.65 dB for TE1"
  • finite-difference time-domain (3D-FDTD) simulations: A numerical method for solving Maxwell’s equations to predict electromagnetic behavior. "designed and verified by three-dimensional finite-difference time-domain (3D-FDTD) simulations across the C band"
  • frequency-division multiplexing (FDM): Parallelization technique that carries multiple signals on distinct RF subcarriers of the same optical carrier. "frequency-division multiplexing (FDM) has provided an electrical-domain route to further parallelization"
  • grating couplers: On-chip diffraction structures that couple light between optical fibers and planar waveguides. "aligned to the TE grating couplers on opposite sides of the chip"
  • guided mode: A discrete electromagnetic field pattern confined within a waveguide. "The vectorial guided-mode dimension of the FieldCore is realized using two guided modes, TE0 and TE1."
  • inductively coupled plasma (ICP) etching: A plasma-based etching process used to transfer nanoscale patterns into semiconductor materials. "electron-beam lithography (EBL) and inductively coupled plasma (ICP) etching steps"
  • insertion loss (IL): The loss of signal power resulting from the insertion of a device in an optical path. "Measured insertion loss (IL) and inter-mode crosstalk (CT) spectra of the test circuit"
  • inter-mode crosstalk (CT): Undesired coupling of power between different guided modes in a multimode system. "Measured insertion loss (IL) and inter-mode crosstalk (CT) spectra of the test circuit"
  • inter-wavelength beating terms: Interference products generated by closely spaced optical wavelengths during detection. "when the inter-wavelength beating terms fall outside the effective electrical bandwidth of the receiver"
  • integrated combs: On-chip multi-wavelength sources that generate many evenly spaced optical lines. "wide-span multi-wavelength sources such as integrated combs52-54"
  • inverse design: Computational photonic design approach that optimizes device geometry to achieve desired optical responses. "Inverse design here enables compact, broadband and fabrication-tolerant implementations of otherwise difficult multi-objective photonic functions"
  • Mach–Zehnder interferometer (MZI): An interferometric photonic circuit used for tunable splitting, combining, and attenuation. "tunable couplers and adders are both implemented using dual-port Mach-Zehnder interferometers (MZIs)"
  • mode-division multiplexing (MDM): Technique that uses multiple guided modes as parallel channels in a single waveguide. "Mode-division multiplexing (MDM) has introduced additional spatial channels without increasing spectral occupancy"
  • mode-insensitive phase shifter (MIPS): A phase shifter designed to impart the same phase shift to multiple guided modes. "mode-insensitive phase shifters (MIPSs), realized by waveguide widening to equalize the thermo-optic response of TE0 and TE1"
  • mode multiplexer/demultiplexer (mode MUX/deMUX): Devices that map signals onto, and separate signals from, different guided modes. "mode MUX/deMUX"
  • mode selectivity: The ability of a component to preferentially excite or process a specific guided mode with minimal crosstalk. "co-optimizing mode selectivity, footprint and bandwidth"
  • multimode-interference coupler: A passive device leveraging multimode interference to combine or split optical signals. "an optical adder implemented with a multimode-interference coupler"
  • noise-aware training: Training methodology that accounts for hardware noise/quantization to maintain model accuracy at lower precision. "A noise-aware-trained46 precision sweep shows that accuracy remains stable down to 4-bit precision"
  • optical carrier: The high-frequency optical wave that is modulated to convey information. "multiplexing RF subcarriers on each optical carrier"
  • peak signal-to-noise ratio (PSNR): A metric quantifying image fidelity by comparing the maximum possible signal to reconstruction error. "The convolution outputs sustain a mode-averaged peak signal-to-noise ratio (PSNR) of 30.5 dB"
  • plasma-enhanced chemical vapor deposition (PECVD): A thin-film deposition technique that uses plasma to lower process temperatures. "an 800-nm-thick SiO2 upper cladding was deposited by plasma-enhanced chemical vapor deposition (PECVD)"
  • photodetector (PD): A device that converts optical signals into electrical signals for measurement. "detected by a 100-GHz photodetector (PD, Finisar XPDV4121R)"
  • radio-frequency (RF) subcarrier: A lower-frequency electrical tone used to carry separate data streams on a common optical wavelength. "multiplexing RF subcarriers on each optical carrier"
  • silicon-on-insulator (SOI): A semiconductor wafer structure with a silicon device layer atop an insulating oxide, used for photonics. "We fabricate a 4 × 4. FieldCore on a silicon-on-insulator (SOI) platform"
  • t-distributed stochastic neighbor embedding (t-SNE): A nonlinear dimensionality reduction technique for visualizing high-dimensional data. "t-SNE visualization of the feature representations produced by FieldCore inference"
  • tera operations per second (TOPS): A measure of computational throughput equal to 1012 operations per second. "an estimated aggregate compute throughput of 69.12 tera operations per second (TOPS)"
  • thermo-optic coefficient: Parameter describing how a material’s refractive index changes with temperature. "yielding a calculated mismatch in the effective thermo-optic coefficient below 1%"
  • titanium nitride (TiN): A resistive heater material used for thermal tuning in photonic circuits. "Integrated heaters were then formed by ... lift-off of 200 nm titanium nitride (TiN)."
  • tunable coupler: A reconfigurable splitter that adjusts the fraction of light coupled into different paths. "Within the FieldCore, tunable couplers and adders are both implemented using dual-port Mach-Zehnder interferometers (MZIs) for programmable splitting and combining"
  • variable attenuator: A device that continuously controls optical power, here realized using an MZI. "the multiplier is realized as a single-port MZI-based variable attenuator for analog weight loading"
  • wavelength-division multiplexing (WDM): Technique that uses multiple optical wavelengths as parallel channels for data. "Wavelength-division multiplexing (WDM) has been widely used to distribute data across many optical carriers"
  • wavelength-insensitive edge couplers: Fiber-chip interfaces designed to maintain coupling efficiency over a broad wavelength range. "broader operating windows enabled by wavelength-insensitive edge couplers"
  • waveguide crossing: A compact structure allowing two waveguides to intersect with low loss and crosstalk. "the waveguide crossing"
View Paper Prompt  View All Prompts 

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Generate Now

Collections

Sign up for free to add this paper to one or more collections.

Sign Up

Tweets

Sign up for free to view the 1 tweet with 68 likes about this paper.

Sign Up for Free