Reconfigurable Compute-In-Memory on Field-Programmable Ferroelectric Diodes.

Area of Science:

• Materials Science
• Computer Engineering
• Solid-State Physics

Background:

• The rise of sensor technology necessitates a shift towards data-centric computing, requiring new device-level innovations for compute-in-memory (CIM) operations.
• Existing CIM architectures face challenges balancing performance and flexibility for diverse data operations.

Purpose of the Study:

• To present a novel transistor-free CIM architecture utilizing Aluminum Scandium Nitride ferroelectric diodes (FeDs).
• To demonstrate the feasibility of integrating FeDs for storage, search, and neural network operations directly onto silicon microprocessors.

Main Methods:

• Development of a transistor-free CIM architecture based on sub-50 nm thick Aluminum Scandium Nitride ferroelectric diodes (FeDs).
• Integration of FeD devices with silicon microprocessors using scalable fabrication processes.
• Leveraging FeD properties like field-programmability, nonvolatility, and nonlinearity for data operations.

Main Results:

• Demonstration of efficient search operations with a cell footprint <0.12 μm² (projected onto 45 nm node technology).
• Successful implementation of neural network operations with 4-bit precision using FeDs.
• FeD devices exhibit multifunctional capabilities for CIM applications.

Conclusions:

• Aluminum Scandium Nitride ferroelectric diodes (FeDs) offer a promising solution for efficient and versatile compute-in-memory platforms.
• The presented transistor-free architecture overcomes key limitations in current CIM designs.
• FeDs pave the way for next-generation data-centric computing architectures.