Photonics Multiply-Accumulation Computations System Based on Residue Arithmetic

Xiaoxuan Ma, Jiaxin Peng, Nicola Peserico, Hamed Dalir, Tarek El‐Ghazawi, Volker J. SorgerACS 2024 · DOI

Abstract

With the explosion of neural network applications and the linked need to perform an enormous amount of mathematical operations, traditional computing systems faced challenges in terms of speed and power consumption. Photonic integrated circuits (PICs), characterized by low latency, low energy consumption, and high bandwidth data processing capabilities, are attracting more and more attention in the advancing neural network research field. However, current photonic solutions require the use of electrical digital-to-analog (DACs) and analog-to-digital (ADCs) converters, which are significantly energy consumption and weaken the inherent advantages of PICs. In this study, we propose and demonstrate a novel approach, a silicon photonics circuit based on residue arithmetic which allows a faster computation of large integer numbers. By employing a specialized routing architecture, the circuit facilitates one-hot data encoding, obviating the need for DACs and ADCs. Experimental evidence substantiates that the power differential between high and low levels exceeds 6 dB in the worst scenario, with an average higher than 10 dB. Furthermore, the introduction of wavelength-division multiplexing can also achieve a 6 dB high/low difference across multiple wavelengths, elucidating the vast potential of PICs for computational applications.
BibTeX
@inproceedings{ma2024photonics,
  author = {Xiaoxuan Ma and Jiaxin Peng and Nicola Peserico and Hamed Dalir and Tarek El‐Ghazawi and Volker J. Sorger},
  title = {{Photonics Multiply-Accumulation Computations System Based on Residue Arithmetic}},
  booktitle = {ACS},
  year = {2024},
  doi = {10.1021/acsphotonics.3c01704},
}

Loading related content...

Figure