NbO_x Mott Memristor-Based Oscillatory P-trit for Ternary Potts Machine

Area of Science:

• Materials Science
• Computer Science
• Physics

Background:

• Probabilistic computing leverages stochastic dynamics for optimization and inference.
• Current hardware often uses binary probabilistic bits (p-bits), limiting efficiency for multi-class problems.

Purpose of the Study:

• To propose and demonstrate a ternary Potts machine (TPM) using probabilistic ternary digit units (p-trits).
• To overcome the limitations of binary systems in probabilistic computing for complex, multi-class problems.

Main Methods:

• Developed a p-trit based on NbOₓ oscillator's electro-thermal dynamics and threshold-switching memristor.
• Implemented a compact circuit to digitize oscillatory signals into two-bit outputs for practical p-trit operation.
• Integrated the p-trit into a microcontroller unit-based TPM prototype for experimental validation.

Main Results:

• Achieved a practical p-trit operation rate of 2.5 µs/bit with 2.4 nJ/bit energy consumption.
• The TPM prototype solved Max-3-Cut problems with ~30% higher accuracy than p-bit systems.
• TPM required ~33% of the computational overhead compared to p-bit computing.
• Simulations showed robust convergence up to 99.96% for weighted ternary number partitioning.

Conclusions:

• The proposed p-trit and TPM offer a more efficient approach for probabilistic computing, especially for multi-class problems.
• Experimental results validate the enhanced accuracy and reduced computational cost of ternary probabilistic computing.
• This work paves the way for advanced hardware implementations in optimization and inference.