Analog optical computer for AI inference and combinatorial optimization

Area of Science:

• Computer Science
• Optical Engineering
• Artificial Intelligence

Background:

• Digital computing's energy demands challenge AI and optimization sustainability.
• Existing unconventional systems often require inefficient digital conversions and face hardware-software mismatches.
• Analog noise is a significant challenge for current analog computing approaches.

Purpose of the Study:

• To introduce a novel analog optical computer (AOC) for accelerating both AI inference and combinatorial optimization.
• To demonstrate a dual-domain computing platform that overcomes limitations of existing systems.
• To showcase a sustainable and efficient computing solution for demanding applications.

Main Methods:

• Developed an analog optical computer integrating analog electronics and 3D optics.
• Implemented a rapid fixed-point search to avoid digital conversions and improve noise robustness.
• Co-designed hardware and a fixed-point abstraction for AI and optimization tasks.

Main Results:

• The AOC accelerates AI inference and combinatorial optimization on a single platform.
• Achieved enhanced noise robustness and efficiency by eliminating digital conversions.
• Demonstrated capabilities in image classification, nonlinear regression, medical image reconstruction, and financial transaction settlement.

Conclusions:

• The analog optical computer offers a promising path for faster and sustainable computing.
• Native support for iterative, compute-intensive models enables a scalable analog platform for AI and optimization innovation.
• Co-design of hardware and abstraction is key to advancing computing technologies.