Van der Waals Ferroelectric CuInP₂S₆-based Multi-slope In-memory Probabilistic Computing.

Area of Science:

• Materials Science
• Computer Engineering
• Quantum Computing

Background:

• Conventional probabilistic computing (p-computing) architectures suffer from a memory bottleneck due to the physical separation of bit generation and storage.
• Probabilistic bits (p-bits) are essential for p-computing, but current designs face limitations in efficiency and integration.

Purpose of the Study:

• To experimentally integrate voltage-tunable stochastic bit generation and non-volatile memory functionalities into a single in-memory device.
• To realize an efficient p-bit using the van der Waals ferroelectric CuInP2S6 (CIPS) material.
• To demonstrate the advantages of in-memory p-computing over conventional architectures.

Main Methods:

• Utilized the stochastic displacement of Cu+ ions and remanent polarization in CIPS under an external electric field.
• Developed an in-memory device combining p-bit generation and non-volatile memory.
• Performed NP-hard simulations to compare in-memory p-computing with conventional p-computing.

Main Results:

• Achieved stable random bit retention (>1000 s) with low power consumption (∼75 nW).
• Demonstrated reduced time-complexity from O(n^2) to O(n^1.5) for NP-hard problems via in-memory p-computing.
• Showcased dynamic tuning of the probabilistic output's sigmoid slope by varying CIPS layer thickness, enabling adaptive control and reducing convergence steps by 400-fold.

Conclusions:

• The CIPS-based in-memory p-bit eliminates data transfer bottlenecks, enabling efficient and high-performance p-computing.
• This integrated device offers a compact, energy-efficient platform for scalable and adaptive p-computing.
• Dynamic sigmoid slope tunability provides adaptive control crucial for optimizing computational tasks.