Multi-state MRAM cells for hardware neuromorphic computing.

Area of Science:

• Spintronics and Neuromorphic Engineering
• Advanced Materials Science

Background:

• Magnetic tunnel junctions (MTJs) are established in sensing and data storage.
• Emerging applications include high-frequency electronics, energy harvesting, and unconventional computing.
• MTJs offer potential for bio-inspired and neuromorphic computing platforms.

Purpose of the Study:

• To design and implement a hardware neural computing device using serially connected MTJs.
• To create multi-state memory cells for programmable quantized weights in neural networks.
• To form an artificial neuron by integrating MTJ multi-cells with CMOS circuitry.

Main Methods:

• Development of a multi-state memory cell using serially connected MTJs.
• Integration of MTJ multi-cells with CMOS-based summing amplifiers and sigmoid function generators.
• Hardware testing of the neural network for handwritten digit recognition.

Main Results:

• The designed neural network achieved handwritten digit recognition with accuracy comparable to software algorithms.
• The system utilized multi-cells composed of four or more MTJs for storing quantized weights.
• The solution demonstrated superior energy efficiency per image processed compared to similar designs.

Conclusions:

• The proposed hardware implementation offers a viable platform for neural computing using MTJs.
• The multi-state memory cell design enables efficient, programmable quantized weights.
• The device shows promise for energy-efficient, bio-inspired computing applications.