Single transistor used to implement neuromorphic behaviour

Researchers from the National University of Singapore have demonstrated that a single, standard silicon transistor can function like a biological neuron and synapse when operated in a specific, unconventional way. Led by associate professor Mario Lanza, the research team’s work presents a scalable and energy-efficient solution for hardware-based artificial neural networks (ANNs). This brings neuromorphic computing — where chips could process information more efficiently, much like the human brain — closer to reality. Their research findings have been published in the journal Nature.

Researchers believe the human brain is, by and large, more energy-efficient than electronic processors, due to the almost 90 billion neurons that form some 100 trillion connections with each other, and synapses that tune their strength over time — a process known as synaptic plasticity, which underpins learning and memory. Scientists have long sought to replicate this efficiency by using artificial neural networks (ANNs).

ANNs have driven advances in artificial intelligence, loosely inspired by how the brain processes information. However, software-based ANNS, such as those powering large language models like ChatGPT, require vast computational resources and electricity. This makes them impractical for many applications. Neuromorphic computing aims to mimic the computing power and energy efficiency of the brain; this requires redesigning the system architecture to carry out memory and computation at the same place — the so-called in-memory computing (IMC) — and developing electronic devices that exploit physical and electronic phenomena capable of replicating how neurons and synapses work. However, current neuromorphic computing systems have been stymied by the need for complicated multi-transistor circuits or emerging materials that are yet to be validated for large-scale manufacturing.

“To enable true neuromorphic computing, where microchips behave like biological neurons and synapses, we need hardware that is both scalable and energy-efficient,” Lanza said.

The NUS researchers have demonstrated that a single, standard silicon transistor, when arranged and operated in a specific way, can replicate both neural firing and synaptic weight changes — the fundamental mechanisms of biological neurons and synapses. This was achieved by adjusting the resistance of the bulk terminal to specific values, which enabled the control of two physical phenomena taking place in the transistor: punch through impact ionisation and charge trapping.

The researchers also built a two-transistor cell capable of operating in neuron or synaptic regime, which they have called ‘Neuro-Synaptic Random Access Memory’, or NS-RAM.

“Other approaches require complex transistor arrays or novel materials with uncertain manufacturability, but our method makes use of commercial CMOS (complementary metal-oxide-semiconductor) technology, the same platform found in modern computer processors and memory microchips. This means it’s scalable, reliable and compatible with existing semiconductor fabrication processes,” Lanza said.

Through experiments, the NS-RAM cell demonstrated low power consumption, maintained stable performance over many cycles of operation and exhibited consistent behaviour across different devices — all of which are necessary for building reliable ANN hardware suitable for real-world applications. The team’s research findings mark a step forward in the development of compact, power-efficient processors that could enable faster, more responsive computing.

Image credit: iStock.com/KENGKAT