A Lab Error That Could Revolutionize Computing
A serendipitous lab discovery could lead to highly energy-efficient neuromorphic chips. Researchers found that a conventional MOSFET with a floating substrate terminal exhibits artificial neuron properties, and by controlling the substrate resistance, a stable neuron-like element can be created. This breakthrough promises to significantly reduce AI power consumption.
In a laboratory in Saudi Arabia in 2024, a student forgot to connect the substrate terminal of a MOSFET while measuring memory, leading to a sharp increase in current with hysteresis resembling the behavior of a biological neuron. Researchers led by Mario Lanza and Steven Pasos found that with a floating substrate, holes generated by impact ionization do not flow to ground but accumulate in the substrate, raising its voltage. When this voltage exceeds a threshold, a parasitic bipolar transistor abruptly turns on, creating a current spike, and then the voltage relaxes. By using an additional transistor to adjust the substrate resistance, the scientists achieved uniform behavior: all neurons fired at the same voltage with low variability. Tests showed stable operation for 10 million cycles without failures. Currently, one neuron is implemented with dozens or hundreds of MOSFETs, while the proposed approach with one or two transistors per neuron could enable scaling systems, approaching the energy efficiency of the brain, which is roughly a million times more efficient than modern neural networks.
Source: IEEE Spectrum AI —
original