LightIN: a versatile silicon-integrated photonic field programmable gate array with an intelligent configuration...

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Integrated Photonics

Background:

• Artificial Intelligence (AI) models demand significant computing power and high-bandwidth communication, challenging conventional electronic circuits.
• Silicon photonics offers high speed, low latency, and large bandwidth, making it suitable for AI data transmission and photonic neural networks.
• Current silicon photonic chips lack adaptability and hardware-software coordination, hindering the development of advanced photo-electronic AI clusters.

Purpose of the Study:

• To develop a reconfigurable silicon photonic chip for multifunctional AI cluster applications.
• To demonstrate diverse functionalities including computing acceleration, signal processing, network switching, and secure encryption on a single chip.
• To establish a framework for testing and controlling the photonic chip without in-chip monitoring.

Main Methods:

• Development of a reconfigurable silicon photonic chip with 40 programmable unit cells and over 160 components.
• Implementation of a custom testing, compilation, and adjustment framework for chip control.
• Demonstration of matrix multiplication, neural network image recognition, wavelength locking, photonic channel switching, and physical unclonable functions.

Main Results:

• Achieved 4x4 bi-direction unitary and 3x3 uni-direction non-unitary matrix multiplications at over 1.92 TOPS with 6.22-bit precision and 1.875 pJ/MAC energy efficiency.
• Demonstrated neural networks for image recognition with a latency of 260 ps and micro-ring modulator wavelength locking in 5-32 Gb/s systems.
• Showcased 4x4 photonic channel switching with -44 dB inter-channel crosstalk and silicon photonic physical unclonable functions.

Conclusions:

• The developed reconfigurable silicon photonic chip integrates diverse AI functions, addressing limitations of current technologies.
• The novel optoelectronic processing system, including the chip and software stack, enables advanced photonic system-on-chip design.
• This work paves the way for constructing efficient and versatile photo-electronic AI clusters.