Universal photonic artificial intelligence acceleration.

Area of Science:

• Photonics
• Artificial Intelligence (AI)
• Deep Learning
• Computer Engineering

Background:

• Photonics research has focused on accelerating tensor operations for AI and deep learning to enhance energy efficiency and performance.
• The field seeks alternatives to traditional computing to sustain progress beyond Moore's Law and Dennard scaling.
• Current photonic chips lack the precision for practical AI, with demonstrations limited to basic tasks.

Purpose of the Study:

• To introduce a novel photonic AI processor capable of executing advanced AI models.
• To demonstrate the potential of photonic computing to compete with electronic AI accelerators.
• To contribute to the development of post-transistor computing technologies.

Main Methods:

• Development of a new photonic AI processor architecture.
• Execution of advanced AI models including ResNet and BERT on the photonic processor.
• Integration and testing of the Atari deep reinforcement learning algorithm.

Main Results:

• The photonic AI processor successfully executed complex AI models and a deep reinforcement learning algorithm.
• The processor achieved near-electronic precision for a variety of workloads.
• This marks a significant advancement in photonic computing for AI applications.

Conclusions:

• The developed photonic AI processor demonstrates practical capabilities for advanced AI tasks.
• Photonic computing is emerging as a viable competitor to established electronic AI accelerators.
• This work represents a crucial step towards future post-transistor computing paradigms.