A sub-10-millisecond neural dynamical system based on phase-change memristors.

Area of Science:

• Hardware acceleration for neural dynamical systems
• Phase-change memristor applications
• Differentiable computing for physical modeling

Background:

• High-fidelity physical-world modeling requires real-time, dense, and differentiable deformation fields.
• Existing Neural Dynamical Systems (NDSs) offer capabilities but suffer from high latency (hundreds of milliseconds).
• Overcoming latency is crucial for practical applications in surface reconstruction and similar tasks.

Purpose of the Study:

• To develop and demonstrate a low-latency hardware implementation of Neural Dynamical Systems (NDSs).
• To leverage phase-change memristors and compute-in-memory for enhanced NDS performance.
• To achieve sub-10-millisecond latency for NDS computations in surface reconstruction.

Main Methods:

• Fabrication of a 40-nanometer NDS chip utilizing phase-change memristors.
• Exploitation of memristor conductance drift and multilevel compute-in-memory capabilities.
• Adaptive stepsize integration with embedded neural networks implemented in hardware.

Main Results:

• Achieved a single-iteration NDS computation latency of 2.12 milliseconds, below the 10-millisecond target.
• Demonstrated 3.82× to 36.27× speed improvement and 11.75× to 24.73× power reduction compared to prior NDS hardware.
• End-to-end NDS latency outperformed GPU A100 by 50.38× to 478.18×.

Conclusions:

• The developed NDS hardware significantly reduces latency and power consumption for complex geometric tasks.
• Phase-change memristors enable efficient, high-performance compute-in-memory for NDS applications.
• This technology paves the way for real-time, high-fidelity physical-world modeling and surface reconstruction.