High-integrated photonic tensor core utilizing high-dimensional lightwave and microwave multidomain multiplexing

Area of Science:

• Photonics and Artificial Intelligence
• Optical Computing Hardware

Background:

• Growing parameter counts in neural networks challenge traditional tensor computing hardware.
• Optical intelligent computing offers a promising alternative to electrical computing, but faces limitations in device size and photonic integration.
• Existing optical chip designs struggle to meet the demands of complex AI computations.

Purpose of the Study:

• To introduce an ultrahigh computing density optical tensor processing unit (OTPU) to address the limitations of current hardware.
• To demonstrate a novel optical tensor core based on microring resonators (MRRs).
• To enable efficient tensor convolution operations using hybrid lightwave and microwave multiplexing.

Main Methods:

• Utilized individual microring resonators (MRRs) as the foundation for the optical tensor processing unit.
• Orchestrated MRR capabilities through independent tuning of multiwavelength lasers to form an optical tensor core.
• Implemented tensor convolution operations via hybrid multiplexing across time, wavelength, and microwave frequency domains.

Main Results:

• Achieved an extraordinary computing density of 34.04 TOPS/mm² with the MRR-based OTPU.
• Demonstrated a high accuracy rate of 96.41% in recognizing MNIST handwritten digits.
• Validated the effectiveness of lightwave and microwave multidomain hybrid multiplexing for tensor operations.

Conclusions:

• The developed MRR-based OTPU represents a significant advancement in optical tensor processing.
• This technology offers a viable solution for high-performance optical chips in artificial intelligence applications.
• The ultrahigh computing density and accuracy pave the way for next-generation AI hardware.