Photonics and microwaves merge to improve computing flexibility.

Area of Science:

• Photonic computing
• Artificial neural networks
• Integrated photonics

Background:

• Artificial neural networks (ANNs) rely on tensor operations, which are computationally intensive.
• Traditional electronic architectures face a storage-and-computing bottleneck, hindering efficient large-scale tensor processing.
• Existing photonic computing solutions often lack the required density and efficiency for complex ANN tasks.

Purpose of the Study:

• To develop a novel photonic tensor processing unit (PTPU) for accelerating ANN computations.
• To overcome the limitations of electronic computing in handling high-dimensional tensor operations.
• To enhance the computing density and efficiency of photonic integrated circuits for AI hardware.

Main Methods:

• A single microring resonator was utilized as the core component for photonic tensor processing.
• Tensor convolution operations were performed by manipulating multiple dimensions: time, wavelength, and microwave frequency.
• Precise control over multi-wavelength lasers enabled the dynamic adjustment of the resonator's operating state.

Main Results:

• The developed PTPU successfully executed multi-dimensional tensor convolution operations.
• A remarkable photonic computing density of 34.04 TOPS/mm² was achieved.
• This density significantly surpasses the performance benchmarks of current photonic computing chips.

Conclusions:

• The microring-resonator-based PTPU offers a promising solution for efficient tensor processing in ANNs.
• This advancement addresses the critical bottleneck in electronic computing for AI acceleration.
• The high computing density achieved paves the way for next-generation, high-performance photonic AI hardware.