Microcomb-enabled parallel self- calibration optical convolution streaming processor

Area of Science:

• Photonics and Neuromorphic Computing
• Data Center Hardware Architectures

Background:

• Exponential growth in cloud computing and AI necessitates high-bandwidth, energy-efficient data center hardware.
• Moore's Law limitations drive the search for alternatives like optical neuromorphic computing for ultra-high speeds and low energy consumption.

Purpose of the Study:

• To propose and validate a microcomb-enabled parallel optical convolution streaming processor (OCSP).
• To demonstrate a practical solution for integrating photonic computing into data center interconnects for AI workloads.

Main Methods:

• Developed an OCSP utilizing time, space, and wavelength three-dimensional multiplexing.
• Implemented a robust self-calibration mechanism for accurate optical phase calibration.
• Employed time-space interleaving passive periodic interference architecture with wavelength-division multiplexing.

Main Results:

• Achieved data rates of 50 GBaud or higher.
• Reached a convolution computing speed of up to 4 trillion operations per second (TOPS).
• Experimentally verified performance for parallel image feature extraction and recognition tasks.

Conclusions:

• The OCSP provides a scalable, low-latency solution for AI workloads in data centers.
• This technology offers a practical pathway for integrating photonic computing units into existing data center infrastructure.