Cornell Researchers Build 20-Nanometer FeMEMS Chip With 200 States for Lower-Power AI
Updated
Updated · Cornell Chronicle · Jun 1
Cornell Researchers Build 20-Nanometer FeMEMS Chip With 200 States for Lower-Power AI
1 articles · Updated · Cornell Chronicle · Jun 1
A Cornell team reported a ferroelectric microelectromechanical device that stores data electrically but reads it through beam vibration, aiming to cut the energy wasted moving data between memory and computation in AI hardware.
The FeMEMS design uses a 20-nanometer hafnium zirconium oxide layer in a suspended beam: electrical pulses set ferroelectric domains, and a small read signal triggers motion that reveals the stored analog value.
Researchers demonstrated about 200 distinguishable electromechanical states, a level of analog precision meant to reduce error buildup when many AI calculations are chained together.
Because the incoming signal interacts directly with the stored state inside the device, the beam’s motion acts as a physical multiplication step, a core operation in neuromorphic and analog in-memory computing.
Next, the group plans larger arrays for matrix operations and integrated control circuitry, while also eyeing uses in adaptive microsystems and ferroelectric-material studies beyond AI.
Why are scientists turning to mechanical motion for computing, a radical departure from decades of purely electronic design?
With AI's energy demand surging, can vibrating microchips avert the coming global power crisis predicted by 2028?
Cornell’s “Microwave Brain” Chip: The World’s First Analog Microwave Neural Network Achieves 88%+ Accuracy with Ultra-Low Power AI
Overview
In August 2025, Cornell University unveiled the groundbreaking 'microwave brain' chip, marking the debut of the first fully integrated microwave neural network on silicon. This innovation represents a major shift in AI hardware, moving beyond traditional digital architectures to deliver unprecedented efficiency and versatility. The chip can handle a wide range of computational tasks, from simple logic operations to complex pattern recognition, all while consuming less energy and space than conventional systems. Its introduction promises to redefine AI capabilities, opening new possibilities for faster, more efficient, and adaptable artificial intelligence solutions.