Thermodynamic Computers Go With the (Energy) Flow

Traditional computing relies on forcing bits to ignore environmental noise to maintain accuracy. Researchers are now exploring a different approach called thermodynamic computing which turns that noise into an actual resource for computation. Instead of fighting thermal fluctuations, these systems harness the movement of energy through circuits to perform calculations, similar to how proteins fold into their stable states within a cell.

Normal Computing, a startup founded by former engineers from Google, has developed prototypes that use networks of electronic circuits to perform complex tasks like matrix inversion. By using electrical resonators that respond to noise, the system reaches a state that encodes the solution to a problem. Because this process runs as the system interacts with its environment, it could significantly reduce the power consumption and heat output that currently plague high-performance computing.

While equilibrium thermodynamic computers wait to settle into a stable state, newer models are designed to compute while moving out of equilibrium. This nonequilibrium approach mimics processes found in nature, where life itself is powered by constant flows of energy. Startups like Extropic are also testing semiconductor-based chips that promise massive gains in energy efficiency for generative AI algorithms compared to existing digital neural networks.

This field is currently in its early stages. Physicists compare the current state of thermodynamic hardware to the early, small-scale quantum computers of the 1990s. While significant hurdles remain before this technology can scale to match modern AI demands, the potential to perform calculations with a fraction of the energy required by current silicon chips has caught the attention of researchers looking for the next phase of computational architecture.