Achieving software-equivalent accuracy for hyperdimensional computing with ferroelectric-based in-memory computing

Area of Science:

• Computer Science
• Materials Science
• Electrical Engineering

Background:

• Hyperdimensional computing (HDC) uses high-dimensional vectors (HVs) for brain-inspired learning, but is memory-intensive.
• Existing in-memory computing (IMC) for HDC often uses binary precision, limiting accuracy and struggling with large HV associative searches.
• Current IMC memory arrays are size-restricted, hindering direct associative search of large HVs crucial for high accuracy.

Purpose of the Study:

• To develop a multi-bit IMC system for HDC that overcomes the limitations of binary precision.
• To demonstrate improved energy efficiency and reduced latency in HDC through IMC.
• To enable efficient associative search of large HVs within an IMC architecture.

Main Methods:

• Implementation of a multi-bit IMC system using ferroelectric field-effect transistors (FeFETs).
• Experimental demonstration of multi-bit, array-level content-addressable memory (CAM) operations with FeFETs.
• Development of a scalable architecture supporting associative search of large HVs using CAMs.

Main Results:

• Achieved software-equivalent accuracies in HDC tasks.
• Reduced HDC system dimensionality.
• Improved energy consumption by 826x and latency by 30x compared to a GPU baseline.
• Successfully demonstrated multi-bit array-level CAM operations with FeFETs.

Conclusions:

• The proposed multi-bit FeFET-based IMC system offers a viable solution for efficient and accurate HDC.
• The developed CAM-based architecture enables scalable associative search for large HVs.
• Further research will investigate the impact of non-idealities on application-level HDC accuracy.