Energy-efficient stochastic computing with superparamagnetic tunnel junctions.

Area of Science:

• Nanotechnology
• Computer Engineering
• Artificial Intelligence

Background:

• Superparamagnetic tunnel junctions (SMTJs) offer a viable nanotechnology for stochastic computing.
• Generating random bitstreams is crucial for stochastic computing applications.

Purpose of the Study:

• To introduce a novel, energy-efficient digital bitstream generator based on SMTJs.
• To develop a co-designed neural network algorithm and circuit for enhanced energy efficiency.

Main Methods:

• Utilized pre-charge sense amplifiers for digitally programmable bitstream generation.
• Designed a convolutional neural network architecture optimized for energy-efficient SMTJ-based bitstream generators.
• Co-designed the algorithm and circuit, adapting neural network mathematics to device-level energy efficiency.

Main Results:

• Achieved significantly higher energy efficiency compared to existing SMTJ and CMOS-based bitstream generators.
• Demonstrated a novel neural network architecture utilizing true random bitstreams from SMTJs.
• The resulting convolutional neural network achieved 97% performance on MNIST with approximately 150 nJ per inference.

Conclusions:

• The proposed SMTJ-based bitstream generator offers a substantial improvement in energy efficiency for stochastic computing.
• Co-designing algorithms with energy-efficient hardware enables novel and efficient neural network architectures.
• This approach represents a significant advancement in low-power AI hardware, outperforming recent proposals.