Related Experiment Video
Updated: May 12, 2026
10:00
Gradient Echo Quantum Memory in Warm Atomic Vapor
Published on: November 11, 2013
12.8K
2D (NH4)BiI3 enables non-volatile optoelectronic memories for machine learning
Bo Tong1,2, Jiajun Xu1,2, Jinhong Du1,2
1Shenyang National Laboratory for Materials Science, Institute of Metal Research, Chinese Academy of Sciences, Shenyang, China.
Nature Communications
|February 13, 2025
Summary
Researchers developed a new optoelectronic memory using (NH4)BiI3 for efficient machine learning. This innovation enables faster, lower-energy artificial neural networks with high accuracy, outperforming GPUs in specific tasks.
Area of Science:
- Materials Science
- Nanotechnology
- Artificial Intelligence Hardware
Background:
- Machine learning (ML) efficiency is crucial for artificial intelligence (AI).
- Optoelectronic memories are key for integrating optical training with electrical inference.
- Current optoelectronic memories struggle with low-light, short-pulse modulation, hindering ML performance.
Purpose of the Study:
- To develop advanced optoelectronic memory for efficient ML.
- To overcome limitations of existing non-volatile resistive state modulation.
- To enhance training accuracy, speed, and reduce energy consumption in AI hardware.
Main Methods:
- Synthesis of a van der Waals layered photoconductive material, (NH4)BiI3.
- Fabrication of an optical floating gate transistor using (NH4)BiI3 as the photosensitive gate.
- Construction of one-transistor-one-memory device arrays.
Main Results:
- Demonstrated adjustable synaptic weights under ultra-dim light without gate voltage.
- Achieved ultra-low training energy consumption and a high number of non-volatile resistive states.
- Device arrays reached ~99% accuracy in Artificial Neural Networks (ANNs).
- Matched GPU performance in YOLOv8 with significantly reduced energy usage.
Conclusions:
- The (NH4)BiI3-based optical floating gate transistor offers superior performance for optoelectronic memories.
- This technology enables highly efficient and accurate ANNs and AI hardware.
- Significant potential for reducing energy consumption in AI computations.
