OptiLookUp: An Optical ROM-Based Loop up Table Engine for Photonic Accelerators

Abstract: Read-only memory (ROM) provides deterministic access to predefined data mappings. Extending ROM concepts to the optical domain enables high-bandwidth, low-latency, and parallel memory access, but realizing compact and reconfigurable optical ROM remains challenging due to loss, wavelength control, and integration constraints. This work presents a high-speed, reconfigurable photonic ROM architecture implemented using integrated microring resonators (MRRs). The ROM encodes predefined input-output mappings directly in the spectral response of the photonic devices, enabling deterministic lookup-based operation without dynamic computation during readout. To improve scalability and reduce cumulative insertion loss, the architecture employs compact banked sub-arrays that are selectively addressed through an optical decoding mechanism. Reconfigurability is achieved using transistor-based optical selectors, allowing different ROM banks to be activated without physical light rerouting or interferometric structures. The proposed photonic ROM is designed and evaluated using device-level simulations based on the GlobalFoundries 45SPCLO silicon photonics platform. Simulation results demonstrate reliable operation at data rates up to 12.5 GHz, with stable light-to-current transfer characteristics obtained through integrated photodiode readout. The optical ROM can be used to implement nonlinear activation functions utilised in photonic accelerator architectures, including sigmoid, tanh, ReLU, and exponential mappings.

• The paper presents a novel optical ROM-based lookup table engine using microring resonators for precise nonlinear function mapping in photonic accelerators.
• It employs a scalable banked sub-array organization with transistor-gated optical selectors, achieving sub-100 ps response times and energy efficiencies in the femtojoule range.
• The reconfigurable REST encoding supports dual digital and analog outputs, enabling efficient on-chip nonlinear activation in photonic neural networks.

OptiLookUp: An Optical ROM-Based Lookup Table Engine for Photonic Accelerators

Introduction

OptiLookUp represents a significant advancement in high-speed, reconfigurable photonic read-only memory (ROM) tailored for photonic accelerator systems. The work systematically addresses persistent limitations of prior optical ROM designs, introducing an integrated microring resonator (MRR) architecture capable of precise nonlinear function mapping, analog and digital dual-mode operation, and banked scalability. Through banked sub-array organization, transistor-gated optical selectors, and device-level design for monolithic silicon photonic processes, OptiLookUp provides deterministic, low-latency lookup operations—critical for efficient photonic implementation of activation functions within deep neural networks. The architecture bridges essential gaps in photonic hardware for AI, especially in nonlinear function realization, enabling new pathways for fully integrated photonic computing pipelines and high-bandwidth, energy-efficient function evaluation without resorting to complex nonlinear photonic effects or feedback topologies.

Architecture and Operation

Integrated Microring ROM and Digital/Analog Readout

The base ROM is realized with cascaded MRRs integrated on silicon photonics. Each microring, depending on its electrical bias, is either resonant or non-resonant at the input optical wavelength, enabling deterministic mapping of input codes to specific optical outputs. The architecture supports both digital and analog operation:

• Digital Mode: Uniform optical power at a single wavelength encodes binary outputs depending on resonance states.
• Analog Mode: Binary-weighted optical powers at distinct wavelengths are parsed via photodiodes to generate an output electrical current proportional to the encoded memory value.

  Figure 1: Schematic of a 4-bit optical ROM using microring resonators, illustrating nonlinear mapping, function-select units, and digital-to-analog mapping.

This dual-mode operation is central: by integrating a balanced photodiode, the architecture achieves immediate digital readout for logic applications and analog current outputs for function evaluation, tightly matching the needs of both digital and analog photonic accelerators.

Scalable Banked Sub-Array Organization

To address both cumulative insertion loss and array scalability, OptiLookUp employs a banked structure. Four banks, addressed by the two most significant bits (MSBs), restrict the optical path per access to five microrings, minimizing loss compared to fully cascaded 16-row prior architectures. Optical selectors, implemented via addressable ring resonators, ensure only the target bank receives power and responds, further suppressing crosstalk and maintaining performance under scaling.

Figure 2: Banked photonic ROM topology shows selective addressing and connection of microring columns via function-select and optical selector circuits.

This approach also decouples the scaling of ROM depth from insertion loss and power penalty, a key limitation in prior microring and SOA-ROMS.

Reconfigurable REST-Encoded Function Implementation

OptiLookUp instantiates a REST (ReLU, Exponential, Sigmoid, Tanh) encoding scheme across the 16-word-line ROM matrix, enabling direct mapping to four canonical activation functions used in neural networks. The FS (function-select) block voltage encodings per word-line define which nonlinear mapping is implemented. This configuration supports reprogramming: by simply modifying FS block encodings, new nonlinearities or activation behaviors can be realized without altering the physical device.

Figure 3: Illustration of the programmable photonic ROM with REST-encoded lookup for ReLU, exponential, sigmoid, and tanh across all banks; includes encoding waveforms and selector circuits.

REST-encoded word-lines can be directly activated by incident optical signals, enabling static implementation of complex nonlinear mappings via physical spectral encoding—a significant improvement over dynamic or feedback-based nonlinear photonic circuits.

Shared Readout Layer and Functional Duality

The outputs from all banks are routed, according to selected operation mode, into a shared readout circuit. For digital operation, a cascaded 2×1 MMI tree delivers waveguide outputs corresponding to digital word lines. For analog readout, the optical powers are merged and translated into an electrical current by a photodiode. The architecture accommodates both discrete-level and continuous-valued function outputs—broadening applicability to both logic-based and analog computing domains.

Figure 4: Block-level schematic of the shared analog/digital readout structure for four-bank photonic ROM, highlighting dual output paths.

Nonlinear Function Realization and Empirical Verification

Simulations on the GlobalFoundries 45SPCLO silicon photonic PDK validate the architecture. Both analog and digitally quantized ROM outputs closely track the ideal nonlinear activation curves for ReLU, exponential, sigmoid, and tanh, with sub-ns response characteristics.

Figure 5: Analog output current response of the optical ROM (red) versus ideal nonlinear function (blue) for ReLU, exponential, sigmoid, and tanh.

Figure 6: Digitally quantized output characteristics from the photonic ROM (red) compared to ideal nonlinear mappings (blue) for the same set of functions.

Notable is the accuracy preserved in the sharp transition and saturation regions (e.g., sigmoid near origin), accomplished through fine-tuned FS block encoding. The adopted mapping covers the entire range $x = -16$ to $15$ with $\Delta x = 2$ resolution per word line, offering a practical tradeoff between resource cost and functional approximation.

Speed, Energy, and Transient Evaluation

Transient simulations for the ReLU function indicate settling times under 80 ps, with output current stabilizing rapidly upon word-line and bank transition—a 12.5 GHz data rate ceiling. Power simulations yield access energies of 259.42 fJ/operation and device-level energies of 14.74 fJ/device, on par or better than extant photonic and high-efficiency electronic ROM implementations.

Figure 7: Measured transient response showcasing effective selector action, bank enabling, word-line activation, and electrical output current dynamics for the ReLU mapping.

Strong claims include both analog and digital duality in output and reconfigurability from a single physical layout—neither realized in prior works. For the tanh mapping, the architecture's current constraints preclude signed output, resulting in a positive-offset curve, highlighted as a concrete limitation.

Comparison with State-of-the-Art

OptiLookUp advances the state-of-the-art on several fronts:

• Scalability: The banked organization decouples ROM size from insertion loss.
• Output Duality: Analog and digital operation modes serve broader compute modalities.
• Speed/Energy: Sub-100 ps operation, fJ/operation energy, and GHz-class throughput.
• Reconfigurability: Lookup table contents (i.e., function mapping) are programmable via FS block encoding and REST word-line selection.
• Generalizability: While validated on neural net activation functions, the mapping mechanism extends naturally to any bounded nonlinear behavior, providing a universal LUT approach for photonic systems.

Compared to previous binary-output-only or static/irreconfigurable photonic ROMs, OptiLookUp achieves higher flexibility, lower latency, and competitive or superior energy efficiency while directly supporting functions critical for AI acceleration.

Practical and Theoretical Implications

Practically, OptiLookUp enables hardware realization of on-chip nonlinear activation without fallback to slow, lossy, or power-intensive electrical or feedback photonic circuits. This directly impacts design of photonic neural networks, signal processors, and other systems that require fast, deterministic LUT-based nonlinearity. Theoretically, encoding nonlinear transfer characteristics statically within spectral device states—instead of physical nonlinear effects or time-evolutionary processes—reframes photonic computation towards deterministic, programmable, and high-bandwidth paradigms. The banked, selector-enabled organization also provides architectural insights transferable to future integrated silicon photonic memories and logic units.

Future directions include further scaling of bank sizes, addition of higher resolution word lines and/or wavelength channels, implementation of alternative nonlinearities, and deployment within fully integrated photonic deep neural network backends. Advanced optical decoders may further reduce input overhead and streamline scaling to larger LUT depths.

Conclusion

OptiLookUp establishes a flexible, high-speed, and energy-efficient photonic ROM leveraging banked microring arrays and programmable selector circuits. Through both analog and digital output duality, REST-encoded nonlinear function realization, and device-level validation on silicon photonic foundry processes, the architecture fills critical gaps in photonic computation—especially for neural network and signal processing pipelines. Its deterministic, LUT-based, and reconfigurable approach sets a new benchmark for optical memory and function evaluation engines, with relevant implications for future AI accelerators, programmable photonic hardware, and compact lookup-based function modules.