Related Experiment Video
Updated: Aug 9, 2026
09:43
Transmission of Multiple Signals through an Optical Fiber Using Wavefront Shaping
Published on: March 20, 2017
10.3K
Time-space-wavelength multiplexed photonic tensor core using WDM SiGe EAM array chiplets
Optics Express
|September 23, 2025
Summary
Photonic neural networks (PNNs) offer energy-efficient AI acceleration. This study demonstrates an 8x8 photonic integrated tensor core (PITC) architecture for AI workloads, achieving high classification accuracy on benchmark datasets.
Area of Science:
- Photonics
- Artificial Intelligence
- Integrated Circuits
Background:
- Photonic neural networks (PNNs) are emerging as a next-generation AI platform.
- They offer significant advances in compute energy and area efficiency by utilizing light's parallelism.
- Leveraging light across space, wavelength, and time is key to PNNs' potential.
Purpose of the Study:
- To present a novel photonic integrated tensor core (PITC) architecture.
- To validate the PITC architecture experimentally for AI benchmark workloads.
- To demonstrate multi-wavelength data and weight encoding in a chiplet-based silicon photonic setup.
Main Methods:
- Developed an 8x8 PITC architecture combining an AWGR-enabled PNN with a SiGe EAM array.
- Utilized a chiplet-based silicon photonic accelerator setup.
- Experimentally validated the architecture using AI benchmark workloads at 20 Gbaud.
Main Results:
- Successfully classified IRIS and MNIST datasets using the PITC architecture.
- Achieved Cohen's kappa-scores of 0.8438 for IRIS and 0.7421 for MNIST.
- Demonstrated the feasibility of the proposed PITC architecture for AI applications.
Conclusions:
- The PITC architecture shows promise for efficient AI acceleration.
- Experimental validation confirms the potential of integrated photonics for AI.
- This work paves the way for advanced photonic AI hardware.
