8-bit states in 2D floating-gate memories using gate-injection mode for large-scale convolutional neural networks

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Science

Background:

• The rapid advancement of artificial intelligence necessitates efficient neuromorphic computing hardware.
• Existing two-dimensional floating-gate memories face limitations in state numbers and stability, restricting their practical application.

Purpose of the Study:

• To introduce gate-injection-mode two-dimensional floating-gate memories as a viable solution for large-scale neural network accelerators.
• To enhance memory performance for neuromorphic computing applications.

Main Methods:

• Utilized a coplanar device structure design.
• Implemented a bi-pulse state programming strategy to achieve multi-bit storage.
• Fabricated and tested an array of 256 devices to assess performance and yield.

Main Results:

• Achieved 8-bit states with intervals greater than three times the standard deviation at 3V.
• Demonstrated stability exceeding 10,000 seconds and cycling endurance over 10^5.
• Attained a device yield of 94.9% for the 256 fabricated devices.
• Successfully performed experimental image convolutions and kernel transplanting on a 9x2 array, with results aligning with simulations.
• Showcased that 8-bit precision fixed-point neural networks achieve inference accuracies close to ideal values.

Conclusions:

• Gate-injection-mode two-dimensional floating-gate memories show significant potential for high-efficiency neuromorphic computing hardware.
• The demonstrated performance metrics and successful integration validate their suitability for neural network accelerators.