Ultra-low power neuromorphic hardware show promise for energy-efficient AI computation

A team including researchers from Seoul National University College of Engineering has developed neuromorphic hardware capable of performing artificial intelligence (AI) computations with ultra-low power consumption. The research, published in the journal Nature Nanotechnology, addresses fundamental issues in existing intelligent semiconductor materials and devices while demonstrating potential for array-level technology.

Currently, vast amounts of power are consumed in parallel computing for processing big data in various fields such as the Internet of Things (IoT), user data analytics, generative AI, large language models (LLM), and autonomous driving. However, the conventional silicon-based CMOS semiconductor computing used for parallel computation faces problems such as high energy consumption, slower memory and processor speeds, and the physical limitations of high-density processes. This results in energy and carbon emission issues, despite AI's positive contributions to daily life.

To address these challenges, it's necessary to overcome the limitations of digital-based Von Neumann architecture computing. As such, the development of next-generation intelligent semiconductor-based neuromorphic hardware that mimics the working principles of the human brain has emerged as a critical task.

The human brain consists of approximately 100 billion neurons and 100 trillion synaptic connections. Synapses store interrelated information through synaptic weights and perform computations and reasoning, serving as the basic units of intelligence.

Neuromorphic hardware based on intelligent semiconductor devices that mimic the brain's synaptic operations relies on memristor devices capable of storing multiple resistance states, leveraging those weights for computation. However, the widely researched amorphous metal oxides used for memristors operate via conductive filaments, leading to charge accumulation in only specific areas. This results in asymmetric and nonlinear synaptic weight adjustments, which leads to inaccuracies in parallel computation and low energy efficiency.

To tackle this issue, Dr. Seung Ju Kim and Professor Ho Won Jang focused on the high ion mobility of halide perovskite materials, which had been attracting attention as materials for next-generation solar cells and LEDs. They concentrated on developing neuromorphic devices based on hybrid organic-inorganic materials. The research team discovered that in newly designed two-dimensional perovskite materials, ions can be uniformly distributed across the surface of the semiconductor.

This breakthrough enabled the successful implementation of ultra-linear and symmetric synaptic weight control, which was previously unachievable with conventional intelligent semiconductors. The theoretical principles of this mechanism were proven through first-principles calculations conducted by a team at POSTECH.

By leveraging the performance of the developed device, the researchers evaluated the accuracy of AI computations performed in hardware. They confirmed that not only with small datasets such as MNIST and CIFAR, but also with large datasets like ImageNET, the device could perform inference with a remarkably small error margin of less than 0.08% within theoretical limits.

Furthermore, through collaborative research with the University of Southern California, it was demonstrated that AI computations could be accelerated with ultra-low power consumption, not only at the device level but also at the array level.

This research, which significantly enhances the energy efficiency of intelligent semiconductor materials and devices, is expected to greatly contribute to reducing the overall energy consumption in AI computation. Additionally, by enabling ultra-linear and symmetric synaptic weight control, it can significantly improve AI computation accuracy and has the potential for application in various fields such as autonomous driving and medical diagnosis. Moreover, this technology is anticipated to spur advancements in next-generation AI hardware technologies as well as innovations in the semiconductor industry.

The technology developed in this study is an upgraded version of the technology presented three years ago in a highlighted paper published by Dr. Kim and Prof. Jang in the journal Materials Today. Patent applications are currently under review both in South Korea and the United States.

Prof. Jang, who led the research, commented, "This study provides crucial foundational data for solving the fundamental problems of next-generation intelligent semiconductor devices. The significance lies in demonstrating that uniform ion movement across the surface of the material is more important for developing high-performance neuromorphic hardware than creating localized filaments in semiconductor materials."

Provided by Seoul National University College of Engineering