High-clockrate free-space optical in-memory computing.

Area of Science:

• Optoelectronics
• Optical Computing
• Artificial Intelligence Hardware

Background:

• Real-time data processing is crucial for autonomous systems and robotics.
• Current edge devices struggle with the power and scalability demands of deep neural networks (DNNs).

Purpose of the Study:

• To introduce a novel fanout spatial time-of-flight optical neural network (FAST-ONN).
• To demonstrate an energy-efficient and scalable hardware solution for DNNs at the edge.

Main Methods:

• Utilized vertical-cavity surface-emitting lasers (VCSELs) for high-speed input modulation.
• Employed high-pixel-count spatial light modulators for in-memory weighting.
• Implemented a 3D optical system with parallel differential readout for signed weight inference.

Main Results:

• Achieved billions of convolutions per second with ultralow latency and power consumption.
• Benchmarked performance with You-Only-Look-Once (YOLO) feature extraction at 100 million frames per second (MFPS).
• Demonstrated in-system backward propagation training with photonic reprogrammability.

Conclusions:

• FAST-ONN offers a scalable solution for energy-efficient DNN deployment on edge devices.
• VCSEL transmitters can enhance free-space optical computing clock rates beyond gigahertz.
• The technology opens new possibilities for scaling up free-space computing hardware.