Scientists Turn Standard Transistor Into Artificial Neuron
Researchers from the National University of Singapore (NUS) have achieved a milestone in artificial intelligence hardware. By operating a standard silicon transistor in an unconventional way, the team has enabled it to mimic both the neuron and synapse, the two fundamental components of the human brain. This finding could drastically reshape the future of neuromorphic computing, the field focused on replicating the brain’s information processing abilities.
Traditionally, artificial neural networks (ANNs) have been modeled after biological brains but only superficially. While they have powered remarkable AI systems, such as ChatGPT, they consume enormous computational resources. In contrast, the human brain remains a model of efficiency, with its 90 billion neurons forming about 100 trillion adaptable connections. This remarkable connectivity, governed by synaptic plasticity, allows humans to learn and remember with minimal energy. Scientists have long aspired to match that performance in machines.
Until now, efforts to build brain-like computing systems faced significant hurdles. Most required either complex arrangements of multiple transistors or new, unproven materials. These challenges have made it difficult to scale the technology or apply it to practical scenarios.
Silicon Transistor, New Potential
Led by Associate Professor Mario Lanza, the NUS team discovered that a single, commonly used silicon transistor could serve both neural and synaptic functions. By carefully adjusting the transistor’s bulk terminal resistance, the researchers harnessed two key phenomena—punch-through impact ionization and charge trapping. These effects enabled the transistor to mimic both neural firing and the modulation of synaptic strength.
Moreover, the team developed a two-transistor cell capable of switching between neuron and synapse modes. They call it “Neuro-Synaptic Random Access Memory,” or NS-RAM. This dual capability is critical for neuromorphic computing, which demands that processing and memory functions happen simultaneously, just as they do in the brain.
Scalable, Reliable, and Efficient
Unlike other approaches that rely on exotic materials or intricate designs, the NUS method uses standard CMOS technology. This is the same manufacturing platform already used in most modern electronic devices. Because of this, it offers an immediate advantage: compatibility with existing semiconductor infrastructure.
In extensive testing, the NS-RAM cells showed low power consumption and reliable performance across repeated use. Their consistent behavior, even across different devices, underscores their potential for real-world AI systems.
As a result, this breakthrough may pave the way for a new generation of compact, power-efficient AI processors. These processors could ultimately make smart devices faster, more capable, and significantly more energy-efficient, bringing us closer to machines that truly think like the human brain.
