A flexible digital compute-in-memory chip for edge intelligence.

Area of Science:

• Flexible electronics
• Artificial intelligence
• Integrated circuits

Background:

• Flexible electronics and AI can revolutionize robotics, wearables, and healthcare.
• Developing flexible hardware for efficient neural network inference is challenging.

Purpose of the Study:

• To present FLEXI, a flexible digital AI integrated circuit.
• To address the challenge of efficient flexible computing for AI tasks.

Main Methods:

• Process-circuit-algorithm co-optimization.
• Digital dynamically reconfigurable compute-in-memory architecture.
• One-shot on-chip neural network deployment.

Main Results:

• FLEXI operates at 12.5 MHz with 2.52 mW power consumption.
• Achieved sub-dollar cost, 70-92% yield, 10^10 error-free multiplications, and >40,000 bending cycles.
• Demonstrated 99.2% accuracy in arrhythmia detection and 97.4% in activity monitoring.

Conclusions:

• FLEXI offers a robust, low-cost solution for flexible AI hardware.
• Enables efficient on-device AI for healthcare and wearable applications.
• Highlights the potential of co-optimization and reconfigurable architectures.