Researchers have identified a method to create artificial neurons and synapses using standard metal-oxide-semiconductor field-effect transistors (MOSFETs), a discovery that could significantly reduce the energy consumption of artificial intelligence hardware. By repurposing the often-overlooked bulk terminal of a common transistor, engineers have demonstrated a way to mimic biological neural behavior in a single, industry-standard device, potentially offering a more efficient alternative to the power-hungry graphics processing units (GPUs) currently used in data centers.
The current reliance on GPUs to power large language models and other AI systems comes with a significant environmental and operational cost. According to industry data, modern GPUs can consume up to 1,000 watts per unit, operating continuously to process the massive datasets required for neural network calculations. This energy demand stands in stark contrast to the human brain, which performs complex computations with vastly higher efficiency. Neuromorphic engineering has long sought to bridge this gap by designing hardware that replicates the integrate-and-fire behavior of biological neurons, yet previous attempts have often required complex circuits composed of dozens or hundreds of transistors, limiting their scalability and commercial viability.
The Role of the MOSFET Bulk Terminal
The breakthrough centers on a component of the MOSFET that is typically grounded in standard digital logic: the bulk terminal. In a standard n-type or p-type MOSFET, this terminal connects to the silicon substrate. Under normal operation, the device functions as a binary switch, manipulating ones and zeros. However, by leaving the bulk terminal floating or applying specific resistance, researchers found that the transistor exhibits nonlinear, neuron-like behavior. This occurs due to impact ionization, a process where charge carriers collide with silicon atoms, creating extra electrons and holes. When the bulk terminal is not grounded, these holes accumulate, triggering a “hidden” bipolar junction transistor within the MOSFET structure. This results in a sharp current spike, effectively mimicking the action potential of a biological neuron.
This discovery allows for the creation of an artificial neuron from a single transistor, a massive reduction in the hardware footprint compared to traditional CMOS-based neural simulations. The devices, referred to as neurosynaptic random-access memory (NSRAM), have shown reliability over 10 million cycles in laboratory testing. Furthermore, by adjusting the bulk resistance, the researchers confirmed that the same MOSFET architecture can function as an artificial synapse, with stable and adjustable conductance. This dual-function capability enables the construction of simple neural circuits where current spikes are modulated by synaptic weight before triggering a response in a neuron-like transistor.
Implications for Edge AI and Energy Efficiency
The potential for integrating these components into existing manufacturing pipelines is substantial. Because the technology relies on industry-standard MOSFETs, it is compatible with existing silicon fabrication facilities. This compatibility addresses a major hurdle for previous neuromorphic designs, which often required experimental, non-standard materials or complex multi-transistor architectures that struggled with cost and yield issues. By utilizing a single-device approach, the design offers a near-100 percent yield with minimal variability, according to findings from the research team.
Initial applications for this technology are expected to focus on edge-AI tasks. Bringing increased intelligence to battery-powered systems—such as wearable sensors, local audio processing, or health monitoring devices—could benefit significantly from the power savings provided by neuromorphic circuits. While current GPUs remain the standard for massive, centralized AI workloads, the development of single-transistor neurons suggests a path toward more energy-efficient, distributed computing architectures. The researchers note that while the technology is still in the experimental stage, future scaling could eventually challenge the dominance of power-intensive processors in broader AI applications.
Next Steps in Neuromorphic Development
Transitioning from a laboratory discovery to a commercial microchip involves several technical challenges. Future work will require the development of more accurate computer models to simulate the behavior of these single-transistor neurons and synapses at scale. Additionally, engineers must design peripheral circuitry to drive these devices and convert signals for use in real-world computing architectures. Further fabrication rounds are planned to optimize performance and ensure the scalability of the NSRAM architecture. As the field advances, these developments will likely be detailed in future technical filings and peer-reviewed publications within the semiconductor and AI research communities.
This research marks a departure from traditional approaches to hardware acceleration, focusing on the fundamental physics of the silicon transistor to achieve brain-inspired performance. As developers continue to refine these circuits, the industry will be watching to see how quickly these components can be integrated into low-power AI systems. Readers interested in the progress of these neuromorphic architectures can monitor updates from major semiconductor research consortia and forthcoming academic journals for further validation of the technology’s scalability.