the feed MANY MINDED · THE BRIEF
KNOWLEDGE · forward · impact 3/5 · 2026-07-20

Computers that use thermal noise instead of fighting it

Thermodynamic computing aims to turn random heat and fluctuations into a computing resource, promising far lower power if it works.

Thermodynamic computing is an emerging approach that tries to use random thermal noise and fluctuations as a computational resource rather than suppressing them, the way conventional chips do. The Computing Community Consortium held its first conference on the idea in 2019, and researchers have now simulated thermodynamic computation inside standard silicon logic circuits, suggesting the basic concepts hold up in principle.

There are two main flavors. In equilibrium thermodynamic computing, a system is driven through its energy landscape toward the minimum that represents an optimal solution — analogous to a protein folding into its stable shape. In the other, systems are pushed away from equilibrium so that the trajectory itself, following Langevin dynamics, encodes the computation. Named groups in the field include David Sivak at Simon Fraser and Stephen Whitelam at Lawrence Berkeley, alongside startups: Normal Computing, where physicist Patrick Coles is chief scientist, and Extropic, which makes thermodynamic sampling units aimed at running generative AI.

The reason to care is energy. Today's computing, and AI in particular, is power-hungry and hot; a machine that exploits noise instead of paying to fight it could run at very low power and dissipate far less heat. If that holds at useful scale, it would cut both the electricity bill and the cooling burden of computation — which increasingly gate how much compute the world can afford.

The honesty here is in the conditional. The field is nascent, with only a small research community, and the encouraging silicon results are simulations that merely appear to work in principle. The transformative case is explicitly conditional; it holds only if the approach works. This is a direction and a bet, not a product. Watch whether the simulations translate into physical hardware that beats conventional chips on real problems.

Source: Quanta