Thermodynamic Computing Promises Energy-Efficient AI Images – IEEE Spectrum

The June issue of IEEE Spectrum is here!
Heat may be 10 billion times as efficient for randomization
Charles Q. Choi is a contributing editor for IEEE Spectrum.
This is a representation of a coupling pattern between representative hidden units and a visible layer from an independent dynamical trajectory of Whitelam’s trained denoising thermodynamic computer.
Generative AI tools such as DALL-E, Midjourney, and Stable Diffusion create photorealistic images. However, they burn lavish amounts of energy. Now a pair of studies finds that “thermodynamic computing” might generate images using one ten-billionth the energy.
At the heart of many AI image generators are machine learning algorithms known as diffusion models. Programmers feed the models large sets of images to which they gradually add noise until they resemble the static on a poorly tuned analog television. They then train neural networks to reverse this process, enabling diffusion models to generate entirely new images given prompting.
However, the AI digital computations that add noise and then conjure pictures from the static are energy hungry. Now a new technique involving thermodynamic computing might generate images “with a much lower energy cost than current digital hardware can,” says Stephen Whitelam, a staff scientist at Lawrence Berkeley National Laboratory in California.

Using Nature’s Noise
Thermodynamic computing employs physical circuits that change in response to noise, such as that caused by random thermal fluctuations in the environment, to perform low-energy computations. For instance, a prototype chip from New York City–based startup Normal Computing consists of eight resonators connected to one another via special couplers. Programmers use the couplers to build a kind of calculator customized for the problem they want to study. Then they pluck the resonators, which introduce noise into the resonator-coupling network, performing the calculation. After the system reaches equilibrium, the programmers can read the solution in the new configuration of the resonators.

In a 10 January Nature Communications article, Whitelam and a colleague revealed it was possible to create a thermodynamic version of a neural network. This lays the groundwork for image generation via thermodynamic computing.
Whitelam’s new strategy would give a thermodynamic computer a set of images. The technique would then allow those stored pictures to degrade by letting the natural random interactions between the computer’s components run until the couplings linking these components naturally reach a state of equilibrium. Next, the strategy would compute the probability that a thermodynamic computer with a given state of couplings could reverse the decay process. Then it would adjust the values of these couplings to maximize that probability.
In simulations run on conventional computers published 20 January in Physical Review Letters, Whitelam found that this training process can lead to a thermodynamic computer whose settings can generate images of handwritten digits. It could accomplish this without energy-intensive digital neural networks or a noise-generating pseudorandom number generator.
“This research suggests that it’s possible to make hardware to do certain types of machine learning—here, image generation—with considerably lower energy cost than we do at present,” Whitelam says.
Whitelam cautions that thermodynamic computers are currently rudimentary when compared with digital neural networks. “We don’t yet know how to design a thermodynamic computer that would be as good at image generation as, say, DALL-E,” he says. “It will still be necessary to work out how to build the hardware to do this.”
Although he calculates that thermodynamic computers might have a huge advantage over regular computers in terms of energy efficiency, “it will be challenging to build a thermodynamic computer that can enjoy all of that advantage. It’s likely that near-term designs will be something in between that ideal and current digital power levels.”
Charles Q. Choi is a science reporter who contributes regularly to IEEE Spectrum. He has written for Scientific American, The New York Times, Wired, and Science, among others.

source