Brain Inspired Circuits for Computation and Communication

The demand for generative AI and large language models is constantly growing, and traditional CMOS hardware is struggling to keep up with the software. Current high performance data centers require 10-100 MW of power, or enough to power 80,000 residences. In contrast, the human brain performs millions of billions of computations each second using only 20 W of power. Evidently, nature has us beat when designing computational hardware.

Studies of neurons in the brain reveal that they communicate with each other through analog spiking signals. This spiking is caused by the neurons' membranes periodically storing and then discharging ions. In other words, the ion concentration is an internal degree of freedom controlling the membrane's ability to conduct electricity. It turns out that the underlying mathematics can work for a lot of different kinds of devices: temperature, ion concentration, or chemical reactions can all provide the necessary degree of freedom, or state variable, for mimicking spiking in biological neurons.

The BRAIN-C3 lab specializes in characterizing and designing these resistive devices with internal states, known as memristors. We are currently investigating both optimal materials for memristors with thermal state, as well as memristors with totally new types of state. We use these devices to design nonlinear oscillators that can act as artificial neurons, or as basic components in communications systems.