Silicon Photonics for Deep Learning

Updated 28 January 2026

Silicon photonics is an integrated optical platform that uses WDM, microring resonators, and MZIs to perform rapid matrix-vector and convolution operations in deep neural networks.
It leverages state-of-the-art optical primitives like on-chip FFTs and programmable unit cells to achieve orders-of-magnitude improvements in throughput and energy efficiency.
Emerging architectures incorporate in-situ training, optical backpropagation, and hybrid photonic-electronic memory, paving the way for scalable, energy-proportional AI systems.

Silicon photonics for deep learning denotes the use of integrated silicon-based optical circuits to perform key neural network operations—especially matrix-vector and convolution operations—critically accelerating both inference and training phases of deep neural networks (DNNs). By leveraging WDM (wavelength-division multiplexing), high-bandwidth and massive parallelism intrinsic to photonics, silicon photonic neural accelerators have demonstrated orders-of-magnitude improvement in throughput and energy efficiency over conventional electronic architectures. Architectures range from scalable “broadcast-and-weight” microring weight banks and MZI meshes to reconfigurable programmable unit cells, with integration of non-volatile photonic/electronic memory, all-optical nonlinearities, and even in-situ photonic backpropagation. Key platforms facilitate both inference and on-chip training, providing path toward energy-proportional AI—from datacenter clusters to edge and IoT scenarios.

1. Photonic Device Primitives and Silicon Photonics Platforms

State-of-the-art silicon photonic deep learning accelerators employ a set of fundamental optical primitives, integrated in CMOS-compatible technology nodes:

Waveguides (Si, SiN) confine optical signals for routing and fan-out/in. SOI (Silicon-On-Insulator) is the primary substrate due to high index contrast and mature process control, although SiN waveguides offer lower scattering and eliminate two-photon absorption, permitting larger fan-in and improved parallelism for GEMM accelerators (Karempudi et al., 2024).
Microring Resonators (MRRs) function as programmable, per-wavelength intensity modulators or spectral-selective weight cells. Each ring is thermally or electro-optically tuned to impart a weight $w_{ij}$ by setting the resonance relative to the input channel, implementing digital or analog values via the drop-port output (Sunny et al., 2021, Zhang et al., 2021). Ring Q-factors span 2,000–20,000; per-ring insertion losses range from 0.2–2 dB.
Mach–Zehnder Interferometers (MZIs) perform unitary vector rotations, phase shifts, and implement beam-splitting and -combining for mesh and butterfly-style architectures. The tunable phase difference between arms enables amplitude and phase modulation, forming the basis for universal matrix multiplication networks and programmable optical meshes (Zhu et al., 2 Apr 2025, Feng et al., 2021).
Wavelength-Division Multiplexing (WDM) allows encoding high-dimensional vectors by assigning each component to one of 10–100 wavelengths carried simultaneously on a shared waveguide, with crossbar fan-in/fan-out and highly scalable MAC (multiply–accumulate) parallelism (Niekerk et al., 2022, Afifi et al., 2024).
Photodetectors (Ge/Si) realize the final weighted summation step by converting interference or directly summed optical intensities to electrical current, with bandwidths up to 70 GHz and responsivity 0.6–1 A/W.
Thermo-optic and electro-optic tuning is used for static alignment (TO; ~mW/ring, ms-μs scale) and dynamic programming (EO; μW/ring, ns scale) of phase shifters and rings (Sunny et al., 2021, Karempudi et al., 2024).

2. Architecture Classes: Matrix Multiplication and Convolution

2.1. WDM “Broadcast-and-Weight” Crossbars

Non-coherent broadcast-and-weight architectures encode input activations as amplitude-modulated light at individual wavelengths, multiplex the signals into a waveguide, and selectively drop fractions of each wavelength’s power using per-wavelength programmable MRRs. The incoherently summed drop-port signals at each photodetector compute the vector–matrix multiplication $y = Wx$ for each output neuron in O(1) time, regardless of vector length, up to the limit imposed by crosstalk, loss, and WDM grid density (Niekerk et al., 2022, Sunny et al., 2021, Sunny et al., 2021).

2.2. Coherent Meshes and Programmable Unit Cells

Clements/Reck MZI meshes realize general unitary transformations by configuring a network of cascaded 2×2 interferometers with embedded phase-shifters, scalable to $N\times N$ matrices. SVD or block-decompositions are used for non-unitary or dimension-reduced transforms. In reconfigurable processors (e.g., LightIn (Zhu et al., 2 Apr 2025)), topologies support synthesis of arbitrary unitary and non-unitary matrices, neural network fully-connected layers, and convolution via hardware compilation.

2.3. On-Chip Optical FFT and Fourier-Domain Convolution

Silicon photonic architectures for FFTs map the Cooley–Tukey “butterfly” onto log-depth nested MZI interferometer networks. These “Optical FFTs” implement the 1-D or 2-D DFT in constant or near-constant time, and convolution is realized via two FFTs, point-wise multiplication, and an IFFT. The dominating latency is set by the time-of-flight through the passive optical structure (tens–hundreds of ps), yielding convolution rates up to 10^4× higher energy/area efficiency than GPUs for moderate N (George et al., 2017, Ahmed et al., 2020, Cottle et al., 2020).

3. Training Modalities: In-Situ, DFA, and Photonic Backpropagation

Most early silicon photonic neural accelerators targeted inference, but current architectures integrate various forms of on-chip or in-situ training:

In-situ backpropagation is realized by dual-use crossbars or MZI meshes that implement both $Wx$ (forward) and $W^T\delta$ (backward error) passes using the same device network (e.g., via symmetric waveguide routing in MRR arrays) (Tang et al., 2024, Dang et al., 2022, Dang et al., 2021).
Direct Feedback Alignment (DFA) replaces sequential backpropagation with concurrent feedback using fixed random matrices, broadcasting error vectors optically to all hidden layers and calculating gradients in parallel by MRR arrays. This achieves >10 TOPS and pJ/MAC energy efficiency in proof-of-principle systems (Filipovich et al., 2021).
Optics-informed training integrates non-ideal device transfer functions, nonlinearities, quantization, and noise directly in the loss or forward model during DNN training, maximizing hardware accuracy and robustness (Tsakyridis et al., 2023, Feng et al., 2021).
Hybrid photonic–electronic memory (e.g., memristor crossbars) interfaces directly with photonic convolution arrays, supporting analog weight storage and update (LiteCON, BPLight-CNN), yielding both training and inference capability (Dang et al., 2022, Dang et al., 2021).

4. Energy Efficiency, Throughput, and Scaling

Silicon photonic accelerators consistently outperform state-of-the-art electronic (GPU/ASIC/FPGA) platforms in compute density, energy-per-MAC, and raw throughput. For CNNs and transformer models, demonstrated metrics include:

Platform	Throughput (GOPS)	Energy Efficiency (TOPS/W)	Area-Normalized Efficiency
LiteCON (Dang et al., 2022)	90,000–100,000	~1,000	500 GOPS/s/W/mm²
TRON (LLM) (Afifi et al., 2024)	12,900	2,500	—
CrossLight (Sunny et al., 2021)	—	52.6 kFPS/W	—
BPLight-CNN (Dang et al., 2021)	99,500	9,327	44,000 GOPS/s/mm²
WDIPLN (Niekerk et al., 2022)	—	<10 fJ/MAC (extrapolated)	<0.1 mm²/8×8 neuron
SONIC (sparse) (Sunny et al., 2021)	—	4.3×10⁷ FPS/W	sub-pJ/bit

Advantages in energy and throughput scale with WDM channel count (limited by FSR, ring finesse, and crosstalk), EO/TO tuning optimization, and multi-core tiling (Dang et al., 2022, Sunny et al., 2021, Karempudi et al., 2024). For example, SiN photonics offers an order-of-magnitude reduction in waveguide/ring loss compared to SOI, directly enabling larger parallelism per ring array (Karempudi et al., 2024). Systems have demonstrated $>10^4 \times$ better compound FOM (power × area × convolution/s) than current GPUs for modest transform sizes (George et al., 2017).

5. Precision, Nonlinearity, and Photonic Memory

Photonic MACs are fundamentally analog, hence precision is governed by the electrical/optical SNR, quantization of modulators/weights, and photodetector noise:

MRR-based synapses have reached >9-bit precision using dithering-based calibration and crosstalk compensation, supporting deeper networks before accuracy loss (Zhang et al., 2021).
Nonlinearities: All-optical (SOA, saturable absorber, nonlinear microresonator) and hybrid EO implementations are utilized for ReLU, sigmoid, or tanh, varying in speed, loss, and power (Tsakyridis et al., 2023, Dang et al., 2021).
Weight storage leverages volatile (TO, EO) and nonvolatile (memristor, PCM, ferroelectric) implementations, with in-situ update minimizing the analog/digital interface cost (Sunny et al., 2021, Dang et al., 2022).

6. System Integration, Programmability, and Emerging Paradigms

6.1. Reconfigurable Processors and Control

Emerging silicon photonic processors (e.g., LightIn (Zhu et al., 2 Apr 2025)) integrate 2D meshes of programmable unit cells (MZIs), software-adaptive digital twins, LUT-based phase calibrations, and live adjoint feedback to implement arbitrary matrix operations and neural network layers. Automation via hardware-software co-design frameworks enables efficient switching between acceleration, switching, and encryption functions within the same chip. Key innovations include thermal eigenmode decomposition for cross-talk mitigation and closed-loop voltage–phase mapping for high-precision operation.

6.2. Sparsity-Optimized and Structured Networks

Sparse accelerator architectures such as SONIC (Sunny et al., 2021) leverage WDM MACs, dynamic zero-skipping in laser and ring driving, and weight clustering to further reduce power, achieve sparse MAC energy benefits ( $>5\times$ over dense photonic, $>8\times$ over sparse electronic devices). Butterfly-structured blocks (Feng et al., 2021) and subspace neural networks reduce the number of required trainable photonic elements by factors of 7–23× while maintaining near-ideal DNN accuracy after hardware-aware quantization and noise injection.

6.3. Delocalized Inference (“Netcast”)

New paradigms such as Netcast distribute the weight matrix optically from the core network to the edge, so that the client device does not store local weights—eliminating local memory accesses and reducing energy/MAC to the attojoule regime (40 aJ/MAC, $<$ 1 photon/MAC in photon-starved regime with superconducting nanowire detectors), with demonstrated MNIST accuracy $>$ 98.8% over 86 km (Sludds et al., 2022).

7. Current Challenges and Future Directions

Even as silicon photonics delivers transformative performance, system-level bottlenecks remain:

Thermal tuning power: Static TO tuning remains the dominant power draw for large numbers of rings but is being mitigated by hybrid EO/TO phase shifters, athermal ring designs, and WDM parallelism (Sunny et al., 2021).
Fabrication variations and drift: Weight quantization, crosstalk, and process-induced resonance variation necessitate closed-loop calibration, hardware-aware or variation-robust training (Zhang et al., 2021, Tsakyridis et al., 2023).
Analog/digital interfacing: The laser/ADC/DAC overhead, electronic bottlenecks in activating photonic nonlinearity, and system-level latency discrepancies are being addressed by integrating high-bandwidth analog-to-digital and digital-to-analog interfaces, co-packaged electronics, and monolithic integration (Sunny et al., 2021, Zhu et al., 2 Apr 2025).
Nonlinearity and memory: High-speed, low-power all-optical nonlinearities and non-volatile photonic memory are still active areas of research, particularly for deep learning applications beyond inference (e.g., recurrent/transformer layers) (Dang et al., 2022, Tsakyridis et al., 2023).
Scaling and integration: Multi-core tile arrays, photonic interconnects, and edge-to-cloud coordination are being enabled via advances in dense WDM, SiN/Si hybrid platforms, and 3D photonic-electronic integration (Karempudi et al., 2024, Zhu et al., 2 Apr 2025).

Future directions focus on multi-band WDM (expanding channel counts), scaling of programmable photonic unit meshes, deploying optics-informed training for physical-layer aware DNN optimization, integrating photonic memory/compute with dense logic in AI data centers, and expanding ultra-low-power inference to IoT and edge scenarios (Sludds et al., 2022, Zhu et al., 2 Apr 2025).

References: