A large scale photonic matrix processor enabled by charge accumulation

Area of Science:

• Neuromorphic photonics
• Artificial intelligence hardware

Background:

• Photonic circuits offer energy and time efficiency for artificial neural networks (ANNs) due to high bandwidth and low loss.
• Scaling current photonic circuits to meet the demands of modern ANNs remains a significant challenge.

Purpose of the Study:

• To address the scaling limitations of existing photonic matrix processors for ANNs.
• To propose and investigate a novel time-multiplexed matrix processing scheme.

Main Methods:

• Overview of matrix sizes in ANNs and comparison with existing photonic matrix processor capabilities.
• Proposal and investigation of a time-multiplexed matrix processing scheme using incoherent optical accumulation.
• Achieved 98.9% accumulation accuracy with 1 ns pulses.

Main Results:

• The proposed scheme virtually increases the size of physical photonic crossbar arrays without electrical post-processing.
• Demonstrated high accumulation accuracy (98.9%) for time-multiplexed optical accumulation.
• Projected capability for all-optical matrix-vector multiplication of 16,000 × 64 matrices on a 51.2 mm² area.

Conclusions:

• The time-multiplexed neuromorphic photonic circuit architecture enables efficient, large-scale matrix operations for ANNs.
• This approach facilitates over 110 trillion multiply-accumulate operations per second, overcoming current scaling challenges.