An ultra-energy-efficient platform for brain-inspired computing

Computers have made significant advances in capabilities and capabilities, surpassing the human brain in tasks such as data processing, predictions and communication. However, the human brain still outperforms computers in energy efficiency.

According to Kaustav Banerjee, a professor of electrical and computer engineering at UC Santa Barbara, computers still consume about 10,000 times more energy than the human brain for specific tasks such as image processing and recognition, despite their superiority in mathematical calculations.

“Making computers more energy efficient is crucial, as the global energy consumption of on-chip electronics ranks fourth in the world’s nation-wide energy consumption rankings, and it is growing exponentially every year, powered by applications such as artificial intelligence.” Furthermore, he added, the problem of computer energy inefficiency is particularly pressing in the context of global warming. “highlighting the urgent need to develop more energy-efficient computing technologies,” Kaustav Banerjee said.

Neuromorphic computing (NM) has emerged as a hopeful solution to address the energy efficiency disparity. By replicating the structure and functions of the human brain, in which processing occurs in parallel across a matrix of low-power neurons, it is possible to achieve energy efficiency comparable to that of the human brain.

The proposal by Banerjee and colleagues, including Arnab Pal, Zichun Chai, Junkai Jiang and Wei Cao, in partnership with researchers Vivek De and Mike Davies of Intel Labs, introduces an ultra-power-efficient platform using transistors 2D transition metal dichalcogenide (TMD)-based tunnel field effect (TFET) system. According to the researchers, this platform has the potential to reduce energy requirements to around 100 times less than those of the human brain.

Neuromorphic computing has been a concept for many years, but recent research has intensified its attention. Circuit developments, enabling smaller, denser transistor arrays, are just the tip of the iceberg in enabling brain-inspired computing with improved processing power and functionality while consuming less power .

The need for a broader range of hardware platforms for neuromorphic computing is highlighted by the multitude of potential applications, including AI and the Internet of Things, which demand high processing capabilities.

To solve this problem, the team introduced 2D tunnel transistors, which emerged from Banerjee’s extensive research into creating high-performance, low-power transistors to meet the growing demand for processing power while maintaining low power consumption. These atomically thin, nanoscale transistors operate efficiently at low voltages and form the core of the researchers’ NM platform, allowing them to mimic the highly energy-efficient functions of the human brain.

2D TFETs feature lower off-state currents and exhibit low subthreshold swing (SS), a measure of a transistor’s ability to efficiently switch from off to on. Banerjee noted that lower SS results in reduced operating voltage and faster, more efficient switching.

Unlike the traditional von Neumann architecture found in today’s computers, where data is processed in sequence and the memory and processing components are separate, consuming continuous power throughout operation, an event-driven system like an NM computer only activates when there is an input to process.

This system also distributes memory and processing functions across an array of transistors. The researchers believe that there is still potential to improve energy efficiency.

“In these systems, most of the energy is lost to leakage currents when the transistors are turned off, rather than during their active state,” Banerjee explained.

A ubiquitous phenomenon in electronics, leakage currents are small amounts of electricity that flow through a circuit even when it is turned off (but still connected to power). According to the article, current NM chips use traditional metal-oxide semiconductor field-effect transistors (MOSFETs), which have high on-state current but also high off-state leakage.

“Since the energy efficiency of these chips is limited by off-state leakage, our approach, which uses tunneling transistors with much lower off-state current, can significantly improve energy efficiency,” » said Banerjee.

When incorporated into a neuromorphic circuit replicating the firing and resetting of neurons, TFETs have demonstrated greater power efficiency than the latest MOSFETs, particularly FinFETs, which use vertical “fins” for greater power control. switching and leaks. Although TFETs are still in the experimental phase, their performance and power efficiency in neuromorphic circuits make them a promising choice for the next era of brain-inspired computing.

According to co-authors Vivek De (Intel Fellow) and Mike Davies (director of Intel’s Neuromorphic Computing Lab), “Once realized, this platform can reduce chip power consumption to two orders of magnitude compared to the human brain – without taking into account interface circuitry and memory storage elements. This represents a significant improvement over what is achievable today.

According to Banerjee, a leading figure in the development of 3D integrated circuits, it may be possible in the future to create three-dimensional versions of these neuromorphic circuits based on 2D-TFET, allowing even more precise simulation of the human brain. This visionary insight comes from someone who played a key role in the widespread commercial adoption of 3D integrated circuits.

Journal reference:

1. Arnab Pal, Zichun Chai, Junkai Jiang, Wei Cao, Mike Davies, Vivek De, and Kaustav Banerjee. An Ultra-Energy-Efficient Hardware Platform for Neuromorphic Computing Enabled by 2D-TMD FET Tunnels. Nature Communications, 2024; DOI: 10.1038/s41467-024-46397-3