Computers have made significant progress in terms of processing power, data storage, and communication abilities, often surpassing human brains in various tasks. However, human brains maintain superiority in one critical aspect: energy efficiency.
“The most efficient computers currently require about 10,000 times more energy than the human brain for specific tasks like image processing, despite outperforming the brain in mathematical calculations,” explained Professor Kaustav Banerjee from UC Santa Barbara. The increasing energy consumption of on-chip electronics, especially driven by applications like artificial intelligence, highlights the urgent need for more energy-efficient computing technologies to address global energy concerns and climate change.
Neuromorphic (NM) computing emerges as a promising solution to enhance energy efficiency. By imitating the structure and functions of the human brain, which processes information in parallel through low-power neurons, NM computing aims to approach the energy efficiency levels of the brain. A recent study published in Nature Communications introduces a highly energy-efficient platform using 2D transition metal dichalcogenide (TMD)-based tunnel-field-effect transistors (TFETs), envisioned to bring energy requirements closer to those of the human brain.
Advances in Neuromorphic Computing
Neuromorphic computing concepts have been studied for decades but have gained momentum more recently. Innovations in circuitry allowing for denser arrays of transistors facilitate more processing with lower power consumption. This presents an opportunity to develop hardware platforms for brain-inspired computing to cater to the increasing demand from applications like AI and the Internet of Things.
The team’s development of 2D tunnel-transistors, from Banerjee’s ongoing research to create high-performance, low-power transistors, serves as the backbone of their NM platform. These atomically thin transistors operate efficiently at low voltages, mimicking the energy-efficient processes of the human brain. The researchers highlight the lower off-state currents and subthreshold swing (SS) of these 2D TFETs, enhancing operational efficiency with faster switching capabilities.
Arnab Pal, the lead author, explains that NM computing architectures are designed to function with minimal firing circuits, similar to how neurons in the brain activate only when needed. Unlike traditional computer architectures like von Neumann, which consume continuous power during operations, NM systems activate selectively in response to inputs, distributing memory and processing functions across transistor arrays. Companies like Intel and IBM have introduced brain-inspired platforms with interconnected transistors, leading to significant energy savings.
The researchers note there is still room for improving energy efficiency in these systems. Addressing off-state leakage currents in off circuits is crucial, as most energy is lost through them rather than during active states. By leveraging tunnel-field-effect transistors with lower off-state currents, they aim to significantly enhance power efficiency compared to current metal-oxide-semiconductor field-effect transistors (MOSFETs) used in NM chips.
Integration of TFETs in neuromorphic circuits showcases superior energy efficiency compared to existing MOSFET designs like FinFETs. While TFETs are still in the experimental stage, their performance and energy efficiency make them a promising candidate for the next generation of brain-inspired computing platforms.
According to Intel’s Vivek De and Mike Davies, this platform has the potential to reduce chip energy consumption to levels comparable to the human brain, representing a substantial advancement in current technology.
Looking ahead, Banerjee envisions the development of three-dimensional versions of these 2D-TFET based neuromorphic circuits to achieve even closer resemblance to the human brain, building upon his contributions to the proliferation of 3D integrated circuits.
