Harnessing the Chaos: The Dawn of Thermodynamic Computing Promises Energy Efficiency and a New Paradigm
The relentless pursuit of accuracy and reliability in computing has long been a battle against noise, the ubiquitous jiggling of atoms driven by thermal energy. This inherent randomness poses a constant threat to the precision required for complex calculations, whether in the familiar classical computers that power our daily lives or the nascent quantum devices poised to revolutionize computation. Traditionally, computer engineers have strived to immunize devices against this environmental hubbub by operating at energy levels far exceeding random fluctuations. However, a groundbreaking field is emerging, one that proposes a radical shift: turning noise from an adversary into an ally, harnessing its power to perform computations.
This nascent discipline, known as thermodynamic computing, aims to design computational systems that leverage the very thermodynamic processes that govern energy distribution and dissipation, processes that inherently increase microscopic randomness. The concept gained formal traction following the Computing Community Consortium’s inaugural conference on thermodynamic computing in 2019. Since then, a dedicated community of researchers has been diligently working to translate theoretical principles into practical applications. Recent simulations have demonstrated the feasibility of thermodynamic computation within standard silicon-based logic circuits, suggesting that the fundamental concepts hold promise in established technological frameworks.
The potential advantages of this approach are profound. Thermodynamic computers could offer unprecedented energy efficiency and drastically reduced heat dissipation, addressing critical limitations of today’s power-hungry devices and the looming threat of thermal meltdown in increasingly dense circuits. Patrick Coles, a physicist at Normal Computing, a New York-based startup at the forefront of this field, articulates the core idea: "The field is about designing computers that exploit thermodynamics as a computational resource." If successful, this paradigm shift could not only reshape the computing industry but fundamentally alter our understanding of computation itself.
The Energy Landscape: Navigating Nature’s Computational Power
The second law of thermodynamics dictates that the entropy of a closed system tends to increase over time, leading to a general trend towards less organized states. This often manifests as energy dissipation through random thermal fluctuations, seemingly rendering it useless for computation. However, certain natural processes have evolved to exploit these very fluctuations, guiding systems towards more organized states.
David Sivak, a statistical physicist at Simon Fraser University, suggests that thermodynamic computing draws inspiration from this natural phenomenon: "I think thermodynamic computing was developed with the thought that it was piggybacking on the computation that ‘already happens out there’ in the rest of the world but [is] not explicitly labeled as such."
In thermodynamics, the evolution of a physical system over time can be visualized as a journey through an "energy landscape." This landscape maps the total energy associated with different configurations of a system’s components. Valleys in this landscape represent stable, low-energy configurations, while peaks signify unstable, high-energy states. A system naturally seeks equilibrium by settling into the deepest energy valley.
A compelling analogy for this process can be found in human digestion. The enzyme lactase, essential for breaking down lactose in milk, achieves its functional shape through a complex folding process. This folding is dictated by the sequence of amino acids that form the lactase molecule. As the chain is synthesized within a cell, thermal fluctuations enable it to explore its energy landscape, ultimately collapsing into its most stable, functional configuration. This natural process is, in essence, a form of thermodynamic computation, solving the intricate problem of protein folding using only the ambient thermal energy within the cell. Once folded into its stable shape within a deep energy well, subsequent thermal fluctuations cause only minor wiggling, preserving the enzyme’s structure and function. This illustrates how natural systems, operating within a noisy thermal environment, can achieve remarkable computational feats with exceptional energy efficiency.
Equilibrium and Nonequilibrium: Two Paths to Thermodynamic Computation
Two primary approaches to thermodynamic computing are being explored: equilibrium and nonequilibrium.
Equilibrium Thermodynamic Computing: This approach mirrors the protein folding analogy. The problem to be solved is encoded into a physical system. Thermodynamics then naturally drives this system through its energy landscape, guiding it towards an energy minimum that represents the optimal solution. Stephen Whitelam, a statistical physicist at Lawrence Berkeley National Laboratory, explains, "You could take a small electrical circuit, let it evolve naturally under its thermally driven dynamics, and then measure its properties once it has attained thermodynamic equilibrium."
Nonequilibrium Thermodynamic Computing: In contrast, this approach involves systems driven away from equilibrium by a continuous energy source, akin to how solar radiation maintains the dynamic state of Earth’s atmosphere. Here, computation occurs as the system traverses its energy landscape. The trajectory of this movement, rather than a final state, encodes the calculation. These nonequilibrium processes are fundamental to many natural phenomena, including life itself, which is sustained by a constant flow of energy and matter.
The dynamics of systems under these nonequilibrium conditions often follow what is known as Langevin dynamics, named after the early 20th-century French physicist Paul Langevin. Langevin dynamics describes a process of relaxation where a system attempts to lower its energy, dissipating energy in the process. However, constant random impulses, such as thermal fluctuations, prevent the system from ever reaching a static equilibrium, continuously pushing it onto new paths.
Whitelam notes that Langevin dynamics can, in principle, be realized in electrical circuits operating at sufficiently low power levels to be influenced by thermal fluctuations. If the energy flow within such a circuit can be designed to correspond to the solution of a computational problem, it can then perform thermodynamic computing.
Both equilibrium and nonequilibrium thermodynamic computers draw energy from thermal fluctuations. However, equilibrium systems require reaching a stable state to complete their computation, a process that can be time-consuming. Nonequilibrium systems, on the other hand, can be engineered to complete calculations within a defined timescale. This inherent speed advantage makes nonequilibrium approaches particularly attractive to researchers like Coles, who states they "can potentially be faster, because you’re not waiting for natural equilibration."
Structure from Noise: Realizing Thermodynamic Computing in Silicon
Current efforts to implement thermodynamic computing predominantly utilize silicon-based circuits as analogous systems. In a significant development last year, Coles and his team at Normal Computing demonstrated that such a circuit could perform complex computations, including matrix inversion – a mathematical operation with wide-ranging applications in fields from machine learning and computer graphics to engineering and finance.
Normal Computing, founded in 2022 by former Google X and Google Brain members Faris Sbahi, Antonio Martinez, and Matthias Tan, unveiled a prototype thermodynamic computer. This device featured a custom-designed printed circuit board comprising eight interconnected clusters of components. Each cluster contained a simple RLC resonator, a combination of a resistor, capacitor, and inductor that generates an oscillating electrical signal at a specific frequency. The circuit is intentionally driven by a random electrical signal, essentially being powered by noise.
The strength of the connections, or couplings, between each pair of RLC resonators can be adjusted, forming a matrix that represents the network’s configuration. This setup can be conceptualized as an electrical analogue of interconnected springs. The core question explored is how the network’s dynamics will evolve when subjected to external stimuli.
The researchers discovered that when the network is "shaken" by noise with an energy comparable to the couplings, the equilibrium fluctuations it undergoes mathematically correspond to the inverse of the coupling matrix. "So you can build your device and come back sometime later and measure its fluctuations, and you’ve done matrix inversion," remarked Whitelam.
While the Normal Computing prototype did not rely on ambient environmental noise (which was too weak to influence the dynamics), researchers introduced artificial noise using a random-number generator, an energy-consuming process. This indicates that such analog models do not yet fully embody the theoretical energy advantages of thermodynamic computing. However, Coles emphasizes that the ultimate efficiency stems from the computation itself running "for free" once initiated by noise. The Normal team’s research suggests that scaling up processing units by increasing the number of nodes in the network could eventually enable them to solve problems faster than conventional digital neural networks, while consuming significantly less energy and dissipating less heat.
Following the unveiling of their analog prototype, Normal Computing has announced its latest thermodynamic computer, the CN101. This new device leverages digital processing on a silicon chip, a more scalable technology than their analog circuits. The CN101 is also capable of performing a broader range of computations, including image generation and molecular simulations. This development is currently undergoing independent expert assessment.
The proof-of-principle work by Normal Computing has inspired further research, including a recent simulation by Whitelam of a nonequilibrium thermodynamic computing circuit. Whitelam’s theoretical model focused on a "denoising" problem. He trained an algorithm on a video of Paul Langevin’s face progressively degraded by artificial noise until it dissolved into static. The simulation then demonstrated the algorithm’s ability to reconstruct Langevin’s face from this static noise.
The simulated thermodynamic computer comprised a network of interconnected nodes, with its state determined by the strength of these connections. During training, the algorithm iteratively adjusted these connections to achieve the configuration most likely to yield an image of Langevin. Whitelam likened this process to a network of springs with adjustable stiffness. He mused, "you could build a thermodynamic computer from real springs," though acknowledging this would likely remain a conceptual curiosity rather than a practical technology.
Whitelam’s research indicated that the resulting dynamics followed a path that dissipated minimal heat, estimating it to be approximately 100 billion times less than would be produced by a digital neural network performing the same task.
The training methodology employed by Whitelam bears resemblance to current techniques used for generative AI algorithms, suggesting that generative AI could be a significant application area for thermodynamic computing. "My algorithm is generative in two ways," Whitelam explained. "First, it turns noise into structure, thereby generating order from disorder. Second, if you train it on a set of images, then it can generate additional images that it hasn’t seen before."
Normal Computing is not the sole entity pursuing this path. Extropic, a Boston-based company also founded in 2022 by a team with experience at Google, IBM, Apple, and Microsoft, announced in October 2025 "the world’s first scalable probabilistic computer." This device, a chip featuring thousands of interconnected semiconductor-based components, is claimed to run generative AI algorithms with an energy consumption approximately 10,000 times lower than existing methods. This work was recently published in the journal npj Unconventional Computing.
Nature’s Blueprint: The Biological Connection
Beyond the potential for low-cost, low-dissipation computing, thermodynamic computing offers intriguing insights into the operation of complex natural systems. Many biological processes, such as signal transduction within cells, can be viewed as forms of information processing. For instance, a signaling molecule binding to a cell surface receptor can trigger a cascade of molecular interactions, ultimately influencing gene expression. Cells exhibit remarkable efficiency in performing these computations, with minimal energy dissipation, largely due to the intrinsic thermodynamics of intermolecular interactions.
This has led to speculation about whether cells themselves function as thermodynamic computers. Whitelam suggests, "To my physicist’s way of thinking, it would be fair to say that nature uses Langevin computers programmed by evolution."
Kunihiko Kaneko, a complex-systems theorist at the Niels Bohr Institute in Copenhagen, finds the core concept of harnessing thermodynamic fluctuations "thought-provoking." However, he acknowledges that "whether this effectively translates to computing in a biological context remains an open question."
The field of thermodynamic computing is still in its nascent stages, comparable to the early days of small-scale quantum computers in the 1990s. Quantum computing has since blossomed into a global industry valued at approximately $12 billion. The technological barriers for thermodynamic computing appear significantly lower; constructing circuits from simple semiconductor-based resonators is considerably more manageable than building quantum bits that require delicate entangled quantum states and often cryogenic cooling. The Normal Computing team posited in their 2025 paper that "The lack of technological barriers for thermodynamic computing can potentially make it a more near-term alternative to quantum computing."
Whitelam draws a similar parallel with the evolution of AI, comparing current thermodynamic computing designs to digital neural networks from around 1990. He believes that, mirroring the trajectory of AI, larger circuits and more extensive training could lead to greater capabilities, though he cautions that "that remains to be seen." If successful, thermodynamic computing promises to inject a new kind of "noise" – innovation and disruption – into the technological landscape.
Correction: July 16, 2026
The Extropic paper describing their scalable probabilistic computer was published in early July 2026. We have updated the reference.
