Abstract
As deep learning models increase in complexity, traditional electronic hardware faces significant bottlenecks in processing speed and power consumption. This paper introduces a silicon photonic accelerator designed to perform high-speed weighted addition, a fundamental operation in neural networks. By employing parallelly cascaded micro-ring resonators (MRRs) on a silicon-on-insulator (SOI) platform, the researchers developed a "weight bank" architecture that leverages the massive bandwidth of photonics while effectively suppressing signal crosstalk. The system demonstrates precise weight configuration through integrated thermal heaters and successful processing of correlated signals, offering a promising path toward scalable, low-latency artificial intelligence hardware.
Introduction
The "post-Moore era" presents a critical challenge for artificial intelligence: traditional electronic processors are struggling to meet the escalating demands for higher throughput and lower energy consumption. Photonics has emerged as a compelling alternative, offering near-instantaneous signal propagation and vast bandwidth through Wavelength Division Multiplexing (WDM). However, implementing complex neural network operations like weighted addition in the optical domain is fraught with technical hurdles, most notably inter-channel and intra-channel crosstalk. This interference limits signal integrity and the number of neurons that can be integrated on a single chip. This work addresses these limitations by proposing a robust MRR-based weight bank architecture.
System Architecture and Methods
The core of the proposed photonic integrated circuit (PIC) is the MRR Weight Bank, fabricated using standard 220nm SOI technology. The architecture integrates multiple functional stages to perform end-to-end weighted summation.
As shown in the system schematic, multiple optical carriers (λ₁ to λ₄) are generated by Distributed Feedback (DFB) lasers and filtered. These carriers are then modulated with input data. The heart of the device is a 4x4 MRR weight bank where micro-rings are cascaded between bus waveguides. Each MRR is independently tunable via integrated N-doped ring-shaped heaters. By applying a control voltage, the resonance of each ring can be thermally shifted, allowing for continuous adjustment of the weight applied to each specific wavelength channel. The final "addition" is performed by on-chip Balanced Photo-Detectors (BPDs), which compute the difference between the "Thru" and "Drop" port signals, effectively executing a weighted sum in the electrical domain.
Experimental Results and Characterization
The fabricated device featured waveguides with a width of 500nm, exhibiting a grating coupler insertion loss of approximately 4.8 dB to 5 dB and minimal waveguide losses (~1 dB). A primary focus of the evaluation was the tunability and signal integrity of the weight bank.
The researchers observed a clear linear relationship between the applied voltage (0V to 4V) and the resonant wavelength shift, measured at approximately 0.016 nm/V. This precise control is essential for accurately mapping neural network weights to the optical domain. Furthermore, the system was tested with partially correlated synthetic signals to validate its signal processing capabilities. As illustrated in the time-domain results, the accelerator successfully produced distinct waveform pairs corresponding to different weight configurations (α = +0.8, 0.0, -0.8), confirming the mathematical validity of the photonic weighted addition model. The use of cascaded second-order MRRs proved vital in suppressing crosstalk, enabling cleaner signal separation compared to standard single-pole filters.
Discussion and Conclusion
While this 4x4 weight bank serves as a successful proof-of-concept, several challenges remain for large-scale deployment. Fabrication sensitivities in the chemical processing stage can lead to performance variations, necessitating redundant calibration circuits. Additionally, scaling the architecture to support the millions of weights required by modern AI models will require careful management of thermal interference between adjacent rings.
Future research will focus on integrating these weight banks with on-chip laser sources and developing non-linear optical activation functions. Such advancements would lead to a fully autonomous photonic deep learning processor capable of outperforming electronic counterparts in both speed and energy efficiency.
Summary updated