Photonic Accelerators: High-Speed Computation

• Photonic accelerators are hardware systems that use optical signals to perform computation with high bandwidth, parallelism, and energy efficiency.
• They employ devices like Mach–Zehnder interferometers, microring resonators, and dielectric laser accelerators to execute multiply–accumulate operations and phase-matched particle acceleration.
• Challenges such as optical losses, calibration drift, and data conversion overhead drive ongoing research into heterogeneous integration and software–hardware co-design.

Photonic accelerators are hardware systems that utilize the properties of light for high-speed, energy-efficient computation, with applications spanning from AI to particle acceleration. By exploiting the high bandwidth, parallelism, and low latency intrinsic to photonic circuits, these accelerators have demonstrated order-of-magnitude improvements in throughput and energy efficiency over purely electronic systems in select domains. A diversity of architectures—ranging from silicon photonic integrated circuits for MAC (multiply–accumulate) operations to on-chip dielectric laser accelerators for particle manipulation—has emerged, each leveraging photonic phenomena for distinct computational tasks.

1. Fundamental Principles and Device Architectures

Photonic accelerators convert information into optical signals—often as multiple wavelengths, amplitudes, or phases—and process these signals through engineered photonic devices. Key device classes include Mach–Zehnder interferometers (MZIs), microring resonators (MRRs), modulator arrays, and photonic crystals. Photonic MAC operations, fundamental to deep neural network (DNN) inference and training, are realized via linear transformations such as

$$\mathbf{Y}_{N\times 1} = \mathbf{W}_{N\times N} \, \mathbf{V}_{N\times 1}$$

where $\mathbf{W}$ encodes weights through phase shifts or resonance tuning, and optical interference or transmission realizes the vector–matrix or matrix–matrix product (Al-Qadasi et al., 2021). These operations take place at speeds dictated by the photonic modulation and detection bandwidth, routinely exceeding tens of gigahertz, far surpassing digital CMOS accelerators.

For particle acceleration, dielectric laser accelerators (DLAs) and photonic crystals achieve field configurations that enable energy transfer from optical fields to charged particles, adhering to strict phase-matching conditions. The multi-channel MIMOSA architecture, for example, uses a silicon photonic crystal with an eigenmode at the $\Gamma$ point ($\mathbf{k} = 0$) to accelerate multiple electron beams, with synchronous optical fields in each channel (Zhao et al., 2020). The field in each channel,

$$E_z^n(x, z) = e^{-j(2\pi z/L)} \left[ A_c^n \cosh\left(\alpha(x-x_n)\right) + A_s^n \sinh\left(\alpha(x-x_n)\right) \right]$$

enables scalable and phase-locked electron acceleration with a figure-of-merit $g \approx 0.51$ in demonstrated designs.

2. System Integration: Hybrid Photonic–Electronic Approaches

A critical transition in the field has been the move from “all-photonic” to hybrid photonic–electronic architectures, mitigating the limitations of optics-only systems. The ADEPT electro-photonic accelerator exemplifies this approach: a photonic “photo-core” implements high-throughput GEMM operations, while a digital ASIC handles nonlinear operations, memory, and dataflow control (Demirkiran et al., 2021). On-chip SRAM buffers store network parameters and activations, and optimized pipelining reduces DRAM stalls and conversion latencies.

The importance of system-level modeling—including data conversion and DRAM overheads—has been underscored by recent architecture-level studies (Andrulis et al., 12 May 2024). These analyses reveal that, despite the energy advantages in the analog–optical domain, cross-domain converters (e.g., DE/AE, AE/AO) and DRAM accesses can dominate system energy unless data reuse and conversion minimization techniques are employed.

Packaging strategies, such as the integration of on-chip modulators, photodetectors, and memory with CMOS logic in a single photonic–electronic IC, further streamline system design. Heterogeneous integration via techniques like photonic wire bonding and advanced multi-chip modules enables scaling and practical deployment (Peserico et al., 2021).

3. Performance Metrics, Scaling Limits, and Trade-offs

Photonic accelerators exhibit high throughput (TOPS–POPS regime) and energy efficiency. For example, silicon photonic MAC circuits can execute matrix–vector operations at speeds $\sim$10–100$\times$ that of digital MACs. However, higher bit-widths, precision, and scaling to large N$x$N networks introduce challenges:

• Energy Efficiency: While photonic MACs operate at 10s–100s fJ/Op, static laser and phase-tuning overheads still cause them to lag behind advanced digital CMOS at high resolution (Al-Qadasi et al., 2021).
• Scaling: Optical losses due to splitting, waveguide attenuation, and coupling restrict the maximum feasible matrix size, with networks above $35\times35$ (MZMs) or $85\times85$ (MRRs) experiencing SNR degradation and precision loss. Innovative use of semiconductor optical amplifiers (SOAs), improved coupling, and low-loss waveguides are active areas for overcoming such barriers.
• Bit-Width and Precision: Dynamic range limitations restrict analog photonic accelerators to 4-bit operations unless architectures like SPOGA are used (Alo et al., 8 Jul 2024). SPOGA’s extended optical-analog dataflow (bit slicing, homodyne summation, in-transduction weighting) enables native INT8 GEMM and achieves up to $14.4\times$ improvement in FPS and $2\times$ improvement in FPS/Watt compared to prior state-of-the-art.

Table: Photonic GEMM Accelerator Scaling (from (Alo et al., 8 Jul 2024))

Architecture	Max Integer Bit-width	Parallelism (core)	FPS Improvement
Conventional	4-bit	High	Baseline
SPOGA	8-bit	High (bit-sliced)	14.4×

4. Reliability, Calibration, and Bayesian/Regularized Learning

Analog photonic circuits are susceptible to drift, noise, and fabrication-induced variability. Several strategies have been developed to ensure robust operation:

• In-situ Calibration: The DOCTOR framework implements dynamic on-chip calibration using adaptive probing and sparse regression to counteract temporally drifting variations (phase errors, temperature drift, crosstalk) in MRR-based tensor accelerators. By monitoring device status and remapping computations to “quiet” tiles, DOCTOR maintains accuracy within 1–5% of ideal operation and reduces recalibration overhead by 2–3 orders of magnitude relative to on-chip training (Lu et al., 5 Mar 2024).
• Bayesian/Regularized Training: Bayesian training frameworks treat phase shifters as random variables and apply regularization or full variational inference. This approach minimizes average phase deviation from passive offsets, reducing both tuning power (>70%) and thermal noise, while providing per-actuator sensitivity for actuator deactivation (up to 31%) to reduce power and complexity (Sarantoglou et al., 2022).
• Nonnegative Transformations: For photonic neuromorphic networks employing incoherent, power-based summation (naturally nonnegative), isomorphic mapping from conventional DNNs to nonnegative networks—together with sign-preserving optimization—enables deployment with minimal accuracy loss and maximally efficient photonic utilization (Kirtas et al., 2023).

5. Applications and Specialized Functionality

AI and DNN Acceleration

Photonic accelerators have demonstrated marked improvements in DNN training and inference workloads. Architectures such as ADEPT realize up to $5.73\times$ higher throughput per Watt compared to electronic systolic arrays, even after accounting for system-level overheads (Demirkiran et al., 2021). Recent dynamically-operated tensor cores (DPTC) in the Lightening-Transformer platform allow truly dynamic operand encoding—critical for attention-based models in LLMs—yielding more than $2.6\times$ energy and $12\times$ latency reductions versus earlier photonic architectures (Zhu et al., 2023). Hybrid approaches like HyAtten, which categorize outputs for low- vs. high-resolution photonic–electronic conversion, provide a $9.8\times$ improvement in performance/area (Li et al., 20 Jan 2025).

Particle Acceleration and Quantum Electron Optics

DLAs using photonic crystals and taper-optimized waveguides (with partial field recycling loops) achieve acceleration gradients above 12 MeV/m and photon utilization exceeding 99% through guided-mode feedback (Li et al., 9 Jan 2025). These platforms implement phase-locked, multi-channel acceleration (MIMOSA) (Zhao et al., 2020), phase-space control via alternating phase focusing (Shiloh et al., 2021), and quantum electron beam shaping for free-electron quantum optics.

Security and Novel Functionality

Reconfigurable photonic meshes serve both as neuromorphic accelerators and physical unclonable functions (PUF) by leveraging process-induced phase randomness as hardware fingerprints. Measured equal error rates in the $10^{-7}$ regime and BER below FEC thresholds demonstrate dual-purpose operation (Sarantoglou et al., 16 May 2025).

6. Challenges, Device Technology, and Future Directions

Although photonic accelerators offer unique advantages, several key obstacles remain:

• Integration and Packaging: The size of photonic components and the limited library of reconfigurable photonic elements constrain integration density. Emerging materials (PCMs, ITO) and advances in co-packaging are expected to drive further scaling (Peserico et al., 2021, Ning et al., 21 Mar 2024).
• Data Conversion and Memory: Cross-domain DAC/ADC conversion and DRAM access frequently dominate energy consumption. Architectural strategies—batching, on-chip buffering, dataflow redesign—are required to realize full-system efficiency (Andrulis et al., 12 May 2024).
• Software–Hardware Co-Design: Future accelerators require EPDA tools and software stacks optimized for optical dataflows, as well as in situ and hardware-aware training protocols to address variability and non-idealities (Ning et al., 21 Mar 2024).
• All-Optical Processing and Nonlinear Operations: Progress in all-optical nonlinear activation functions, multi-operand optical devices, and photonic memory promises further reductions in overhead and latency.

7. Distributed Systems and Disaggregated Memory

Novel platforms such as the Photonic Fabric Appliance (PFA) disaggregate HBM memory and on-module photonic switching, enabling up to 32 TB shared memory and 115 Tbps switching bandwidth per system (Ding et al., 18 Jul 2025). This architecture breaks the fixed memory-to-compute ratio inherent in legacy XPU designs, allowing accelerators to accommodate much larger models and parallel computations. CelestiSim simulation demonstrates up to $7\times$ throughput improvements and up to 90% energy reductions in collective data movement, validating the system-level benefits of photonic fabrics for large-scale AI (Ding et al., 18 Jul 2025).

---

Photonic accelerators have rapidly matured from specialized research prototypes to flexible, high-performance platforms addressing bottlenecks in deep learning, scientific computing, secure hardware authentication, and particle physics. System-level design that fuses photonic and electronic technologies, coupled with robust calibration and novel device integration, is central to advancing this field—pushing the boundaries on speed, energy efficiency, and new physical capabilities.