Artificial neuron on silicon chip discovered; mimics brain efficiency, could slash AI power use

Researchers accidentall discovered a simple, single-device artificial neuron and synapse built into ordinary CMOS transistors, according to a new article in IEEE Spectrum. The finding emerged from studying an ordinary CMOS transistor—not even a particularly good one—and leveraging a fourth terminal (bulk connection) that typically gets little attention. By applying voltage to both the gate and bulk terminals, researchers found the transistor exhibits neural-like behavior without requiring the bulky interconnected capacitor networks that have plagued neuromorphic hardware for decades.

The key breakthrough is that a single transistor can now serve as both an artificial neuron and synapse, eliminating the scaling barrier that has prevented neuromorphic systems from competing with GPUs. Prior approaches required wiring many separate CMOS transistors together to approximate a single neuron, dramatically increasing energy overhead, wiring complexity, and silicon footprint. The new device's I-V curve—how current responds to voltage—mirrors biological neuron action potentials: it exhibits a threshold behavior and self-sustaining dynamics that make it ideal for mimicking neural spiking without external oscillation circuits.

For architects evaluating edge AI, this is a long-term play: the accidental discovery sits at the intersection of solving neuromorphic hardware's biggest bottleneck (scale and density) while using industry-standard fabrication. No new process nodes or exotic materials required. The human brain operates on ~20 watts; modern GPUs can consume kilowatts. If this single-device neuron can scale to millions of neurons per chip using standard CMOS, it could enable true brain-scale compute on milliewatt budgets—but that's years away from production. Watch for follow-up papers on scaling and system integration.