All-analog photoelectronic chip for high-speed vision tasks.

Area of Science:

• Optoelectronics
• Photonic Computing
• Artificial Intelligence Hardware

Background:

• Photonic computing offers potential for faster, energy-efficient vision data processing.
• Existing systems face challenges with optical nonlinearities, high power consumption of analog-to-digital converters (ADCs), and noise vulnerability.

Purpose of the Study:

• To propose and experimentally demonstrate an all-analog chip combining electronic and light computing (ACCEL).
• To overcome the limitations of current photonic computing systems for practical deployment.

Main Methods:

• Developed an all-analog chip (ACCEL) integrating electronic and optical computing components.
• Employed diffractive optical computing for feature extraction, directly utilizing light-induced photocurrents for analog computation.
• Integrated adaptive training for joint optoelectronic computing optimization.

Main Results:

• ACCEL achieved 74.8 peta-operations per second per watt energy efficiency and 4.6 peta-operations per second computing speed (over 99% optical).
• Demonstrated low computing latency of 72 ns per frame without ADCs.
• Achieved competitive classification accuracies (85.5% Fashion-MNIST, 82.0% ImageNet, 92.6% video recognition) with robust performance in low-light conditions.

Conclusions:

• ACCEL represents a significant advancement in photonic computing, offering substantial improvements in speed and energy efficiency.
• The proposed architecture effectively addresses key challenges, paving the way for practical applications in areas like wearable devices and autonomous driving.