Lab Accident That Could Revolutionize Computing

Researchers accidentally discovered that a single MOSFET transistor can mimic a neuron when its bulk terminal is left floating. This finding could lead to neuromorphic chips that are far more energy-efficient than current GPUs used for AI.

While testing memory circuits, a student forgot to connect the bulk terminal of a MOSFET, causing the transistor to exhibit neuron-like electrical behavior: a sudden current spike when voltage reached a threshold, then self-relaxation. The team realized that with the bulk floating, impact ionization-generated holes accumulate in the silicon bulk, turning the transistor into a bipolar junction device that fires spikes similarly to biological neurons. By controlling the bulk contact resistance with a second transistor, they achieved consistent, reliable firing across millions of cycles. This accidental discovery could enable energy-efficient neuromorphic computing using standard CMOS transistors, potentially reducing AI's massive power consumption. Currently, GPUs consume up to 1,000 watts each, and neuromorphic chips using complex transistor circuits have already shown up to 1,000-fold power reduction. A single-transistor neuron would allow even larger, more efficient systems.

Сокращения: CMOS = Complementary Metal-Oxide-Semiconductor — комплементарная структура металл-оксид-полупроводник; GPU = Graphics Processing Unit — графический процессор; MOSFET = Metal-Oxide-Semiconductor Field-Effect Transistor — полевой транзистор со структурой металл-оксид-полупроводник

Source: IEEE Spectrum AI — original