Precision-extension technique for accurate vector-matrix multiplication with a CNT transistor crossbar array.

Area of Science:

• Materials Science
• Computer Engineering
• Nanotechnology

Background:

• Machine learning algorithms heavily rely on multiply-accumulate operations, impacting computation time and energy efficiency.
• Resistive networks in crossbar arrays offer potential acceleration for vector-matrix multiplications via Kirchhoff's current law.
• Challenges in resistive crossbar arrays include device variability and crosstalk, limiting precision for high-performance computing tasks.

Purpose of the Study:

• To experimentally demonstrate a precision-extension technique for carbon nanotube (CNT) transistor crossbar arrays.
• To overcome the limitations of device variability and crosstalk in resistive crossbar computing.
• To enable high-accuracy vector-matrix multiplication for in-memory computing applications.

Main Methods:

• Utilized a 10 × 10 carbon nanotube (CNT) transistor crossbar array.
• Implemented a novel precision-extension technique by distributing bit width across multiple devices.
• Leveraged Kirchhoff's current law for computation within the crossbar architecture.

Main Results:

• Achieved high-accuracy vector-matrix multiplication using the developed CNT transistor crossbar array.
• Demonstrated that distributing bit width across multiple devices effectively extends precision.
• Successfully addressed challenges posed by device variability and sneak path currents.

Conclusions:

• The developed precision-extension technique enhances the accuracy of CNT transistor crossbar arrays for in-memory computing.
• This approach makes in-memory computing a viable and attractive option for high-performance computing environments.
• Experimental validation confirms the potential of CNT-based crossbar arrays for demanding computational tasks.