Compute in-Memory with Non-Volatile Elements for Neural Networks: A Review from a Co-Design Perspective.

Area of Science:

• Computer Science
• Materials Science
• Electrical Engineering

Background:

• Deep learning's ubiquity strains traditional computer architectures.
• The von Neumann bottleneck, caused by data movement between memory and compute units, limits efficiency.
• Compute-in-memory (CIM) presents a promising hardware solution to mitigate this bottleneck.

Purpose of the Study:

• To review cross-bar based CIM for neural networks from a co-design perspective.
• To connect material properties with design constraints, architecture, and performance.
• To assess CIM's potential for both training and inference using digital and analog memory.

Main Methods:

• Review of cross-bar architectures utilizing non-volatile memory elements for analog multiply-and-accumulate operations.
• Analysis of how memory material properties impact system-level characteristics (speed, power, accuracy).
• Evaluation of digital and analog memory technologies for CIM applications in neural networks.

Main Results:

• Memory materials critically influence CIM speed, power consumption, and classification accuracy.
• Cross-bar CIM architectures enable efficient matrix-vector multiplication essential for neural networks.
• The review provides metrics for non-volatile memory material properties required for successful CIM.

Conclusions:

• CIM, particularly cross-bar architectures with optimized non-volatile memory, is crucial for efficient deep learning hardware.
• A co-design approach linking materials, architecture, and applications is vital for advancing CIM technology.
• Further development of non-volatile memory materials is necessary to meet the demands of future CIM systems for neural network acceleration.