Thermodynamic Computing: Noise as the Power Source

Unconventional computing: how a thermodynamic computer uses noise

Berkeley Lab describes thermodynamic computing as a branch of unconventional computing. Unconventional computing, in general, means designs that depart from the standard digital processor. Berkeley Lab says the approach inverts the paradigms of both classical and quantum computing and uses thermal fluctuations as its power source.

Thermal noise, which Berkeley Lab calls thermal fluctuations, is the small random motion that heat causes in matter. In a thermodynamic computer, that motion is the source of power.

The same Berkeley Lab page frames the energy question with an everyday figure. A single Google search consumes enough energy to power a six-watt LED for three minutes.

How a stochastic thermodynamic neural network is trained

Berkeley Lab says a thermodynamic computer is a stochastic system. Stochastic means chance is part of how the system behaves, so no two runs look the same. Berkeley Lab adds that the methods used for training digital neural networks do not apply.

The Nature Communications paper "Nonlinear thermodynamic computing out of equilibrium," by Berkeley Lab's Stephen Whitelam and Corneel Casert, takes up the training question. Its authors simulate a digital model of a thermodynamic neural network.

The authors show that a genetic algorithm, a search method that tries many parameter settings and keeps the better ones, can adjust that network's parameters. The adjusted network then performs nonlinear calculations, where the output does not simply scale with the input, at specified observation times. The network in that paper is a simulated model, not a built device.

Thermodynamic AI and energy efficiency: costly to train, cheap to run

Berkeley Lab states the energy efficiency trade-off in one sentence. Its training framework is considerably more costly than the methods used to train digital networks.

The result, once built and trained, is a computer that can operate using very little energy. The sentence gives no figures for either side.

Casert draws the line between training and inference. Inference means running a trained model to get answers. Training by digital simulation is expensive, he says, but once the network is trained and built as physical hardware, inference on it would cost very little energy.

Berkeley Lab sums up the aim as a design that could drastically reduce the energy requirements of machine learning. Neither the news page nor the paper's abstract puts a number on that reduction, such as watts saved or a ratio against a GPU.

Extropic TSU: a thermodynamic sampling unit built on standard CMOS

Extropic describes its TSUs as chips that use the natural thermodynamic fluctuations of standard CMOS transistors. CMOS is the ordinary process used to make most of today's chips. Extropic gives this description on a page titled "Thermodynamic Computing Chips in America".

The chips use those fluctuations to sample directly from programmable probability distributions in a power-efficient manner, according to Extropic. Sampling means producing random outcomes, each as often as its probability says, and programmable means those probabilities can be set. Extropic calls the capability useful for generative AI and for simulations of biology and markets, among other uses.

Thermodynamic computing hardware: what Extropic has and what is open

Berkeley Lab's design has not been built as hardware yet; Extropic is a separate company that describes chips of its own. The Berkeley Lab results come from digital simulation, and the team is looking for experimental partners to make both hardware and software a reality.

According to Whitelam, it is important to work out how to realize these designs in hardware. Existing algorithms assume systems at equilibrium, so Berkeley Lab says new ones will need to be developed as well. Extropic's TSUs are a different approach.