# In Situ Quantization with Memory‐Transistor Transfer Unit Based on Electrochemical Random‐Access Memory for Edge Applications

## TL;DR

This paper introduces a compact synaptic unit that enables efficient low-precision neural network training and inference at the edge with minimal energy consumption.

## Contribution

A novel synaptic unit combining ionic nonvolatile memory and transistors enables in situ weight quantization without extra programming.

## Key findings

- The synaptic unit achieves classification accuracy comparable to ideal quantization methods in binary neural networks.
- ECRAM- and RRAM-based units show 25.51× and 4.84× energy efficiency improvements over traditional digital platforms.
- The design supports continual learning and mitigates catastrophic forgetting in low-precision computations.

## Abstract

In‐memory computing based on nonvolatile synaptic arrays with computing functions has significantly improved the computing energy efficiency of neural networks. However, current synaptic devices are mostly limited to accelerating matrix‐vector multiplication operators, and the differentiated requirements for device characteristics in the training/inference stage have led to a sharp increase in the integration complexity of hybrid synaptic units. Hence, for low‐precision quantization calculations of networks, a compact synaptic unit based on ionic nonvolatile memory‐transistor coupling integration, which enables in situ approximate weight quantization without additional binary programming while maintaining parallel MVM computing capabilities, is developed. Results show that the quantization function, derived from the cell's physical electrical properties, achieves classification accuracy in binary neural networks comparable to the ideal quantization function. This approach supports low‐precision continual learning, mitigates catastrophic forgetting, and enables efficient computations for binary/ternary large language models. At a 4 Mb array scale, ECRAM‐ and RRAM‐based units achieve energy consumption advantage of 25.51× and 4.84×, respectively, over traditional digital platforms, offering a robust in situ quantization framework for low‐precision edge training.

By combining ionic nonvolatile memories and transistors, this work proposes a compact synaptic unit to enable low‐precision neural network training. The design supports in situ weight quantization without extra programming and achieves accuracy comparable to ideal methods. This work obtains energy consumption advantage of 25.51× (ECRAM) and 4.84× (RRAM) over digital platforms, showing the potential for efficient edge AI and large language model computations.