Integrated Neuromorphic Photonic Computing for AI Acceleration

Abstract

Reviews emerging devices, network architectures, and future paradigms for neuromorphic photonic computing as AI accelerators. Demonstrates that in-memory-computing photonic tensor cores can achieve predicted compute density of 880 TOPS/mm² and efficiency of 5.1 TOPS/W.

Key Contributions

• Photonic tensor cores: 880 TOPS/mm² density, 5.1 TOPS/W efficiency (predicted)
• 1-3 orders of magnitude improvement over digital electronic accelerators in both density and efficiency
• MIT photonic chip: all key deep neural network computations optically on-chip, <0.5ns latency, >92% accuracy
• Monolithic coherent optical neural networks via commercial silicon photonics: 92.5% vowel classification, nanosecond latency, femtojoule efficiency

Results

The photonic approach shows dramatic advantages for matrix-multiply-heavy workloads (the backbone of neural networks). In-situ training enables networks to be trained directly on the photonic hardware. Key milestone: fully integrated photonic processor completing classification in less than half a nanosecond.

Limitations

• Predicted performance numbers not yet achieved at scale
• On-chip nonlinearity remains a challenge for deep architectures
• Manufacturing yield and reproducibility concerns
• Integration with existing electronic infrastructure requires co-design

Source: Neuromorphic Photonic Computing — Advanced Materials