Dynamic pass bias control for temperature-resilient neural networks using vertical NAND flash memory

Area of Science:

• Solid State Physics
• Computer Engineering
• Artificial Intelligence

Background:

• Vertical NAND (V-NAND) flash memory is suitable for neuromorphic computing due to its density and scalability.
• Temperature variations cause V-NAND synaptic weight instability, degrading neural network inference accuracy.
• Existing solutions often require reprogramming or additional hardware, limiting practical application.

Purpose of the Study:

• To propose and validate a dynamic pass bias (DPB) control scheme for V-NAND flash memory.
• To enhance the temperature resilience of V-NAND based neural networks.
• To offer a hardware-efficient solution for mitigating temperature-induced weight variations.

Main Methods:

• Implemented a dynamic pass bias (DPB) control scheme by adaptively adjusting pass bias on unselected word-lines during read operations.
• Developed a passive, temperature-adaptive biasing circuit using a silicon MOSFET and V-NAND strings with opposing thermal characteristics.
• Validated the DPB scheme through experimental measurements on commercial V-NAND devices and simulations using CIFAR-10 image classification with a VGG-11 network.

Main Results:

• Experimental data showed significant bit-line current shifts in V-NAND devices with increasing temperature.
• DPB scheme simulations demonstrated substantial mitigation of accuracy degradation across a wide temperature range.
• Adjusting pass bias at lower temperatures improved classification accuracy by up to 10.5%p compared to fixed-bias operations.

Conclusions:

• Dynamic pass bias (DPB) control, both digital and circuit-assisted, effectively compensates for temperature-induced V-NAND weight variations.
• The proposed DPB scheme offers a lightweight and scalable solution for improving the temperature resilience of V-NAND flash memory in neuromorphic applications.
• This approach enhances the reliability and accuracy of V-NAND based neural networks under thermal stress.