High-precision time-domain parallelism photonic computing

Area of Science:

• Photonics
• Artificial Intelligence
• Computer Engineering

Background:

• Artificial intelligence (AI) compute demands are rapidly increasing, exceeding the capabilities of current electronic hardware, leading to an "AI compute gap."
• Photonic computing offers potential solutions with high speed, low latency, and parallelism, but faces challenges in core array scalability and optoelectronic bandwidth efficiency.

Purpose of the Study:

• To propose and validate a novel time-domain parallelism (TDP) computing architecture for AI workloads.
• To develop and characterize high-performance indium tin oxide (ITO)-based photonic devices for the TDP system.
• To demonstrate the scalability and efficiency of photonic computing for addressing the AI compute gap.

Main Methods:

• Developed a time-domain parallelism (TDP) architecture that temporally multiplexes compact cores for large-scale convolutional operations.
• Fabricated indium tin oxide (ITO)-based photonic devices with record 9-bit dynamic reconfigurability (60 kHz) and a 3-dB bandwidth of 311 kHz.
• Validated the TDP system with a 2x2 ITO photonic array on the Fashion-MNIST dataset.

Main Results:

• ITO photonic devices achieved exceptional endurance (>10^10 cycles) and compact phase shifter lengths (20 micrometers).
• The TDP system achieved 84.91% accuracy on Fashion-MNIST classification, surpassing a 5-bit benchmark and matching the FP32 software baseline.
• The system demonstrated high-fidelity performance at a high operating speed of 1694 frames per second.

Conclusions:

• The proposed time-domain parallelism (TDP) architecture offers a scalable and energy-efficient pathway for photonic computing in AI.
• The developed ITO-based photonic devices enable high-performance, reconfigurable optical components crucial for advanced AI hardware.
• This research bridges the AI compute gap by showcasing the practical viability of photonic solutions for demanding AI workloads.