Thermodynamic Computing Overview

• Thermodynamic Computing is a paradigm that exploits non-equilibrium thermodynamic processes like energy dissipation and fluctuations to encode computational tasks.
• It utilizes diverse physical architectures, such as analog circuits, spin-based devices, and quantum thermal machines, to achieve energy-efficient computation.
• Algorithms in TC leverage noise-driven dynamics for matrix operations, logic functions, and generative modeling, balancing tradeoffs among energy, speed, and accuracy.

Thermodynamic Computing (TC) is a computational paradigm in which information processing is carried out by engineered physical systems whose evolution is dominated or fundamentally shaped by thermodynamic processes—fluctuations, dissipation, and self-organization—in contact with an environment. Computation is realized by tuning energetic or informational potentials so that the hardware adaptively explores its state space, settling into configurations that encode the solutions to computational or logical tasks. Unlike traditional digital computing, TC designs computation directly in the thermodynamic substrate, making energy dissipation, control of entropy flow, and management of fluctuations central to both the theory and practice of information processing (Conte et al., 2019).

1. Fundamental Principles and Theoretical Foundations

TC is grounded in the laws of nonequilibrium statistical physics and information theory. A thermodynamic computer is an open, multi-scale system designed to exploit free-energy dissipation, unavoidable thermal/quantum fluctuations, and adaptive self-organization (Conte et al., 2019). The dynamics obey the first and second laws of thermodynamics, with key constraints such as Landauer's bound: every logically irreversible operation incurs a minimal average heat dissipation of $k_B T \ln 2$ per bit erased.

Fluctuation theorems, including the Jarzynski equality and Crooks relation, extend these constraints to the microscopic, stochastic regime: for any trajectory γ and its time-reversal $\tilde{\gamma}$, the probability ratio connects to entropy production and irreversibility,

$$\frac{P(\gamma)}{P(\tilde{\gamma})} = \exp \left(\frac{Q(\gamma)}{k_B T}\right).$$

These relationships enable TC to precisely account for the thermodynamic cost of computation in both equilibrium and highly nonequilibrium regimes (Conte et al., 2019, Rolandi et al., 7 Jan 2026).

The informational capacity of physical computation is further linked to entropy production by the information-processing second law,

$\langle W \rangle \geq k_B T (h_\mu - h_\mu')$

where $h_\mu$ and $h_\mu'$ are input and output Shannon entropy rates.

Traditional models such as Ising networks, hopfield nets, and Boltzmann machines are subsumed as special cases of thermodynamic neural networks where both network state and connectivity evolve under energy minimization and stochastic kinetics.

2. Elementary Devices and Architectures

TC architectures are realized with a diverse range of physical hardware platforms:

• Continuous-variable or analog systems: RLC circuits (Melanson et al., 2023), coupled mechanical oscillators, or Josephson junction arrays, in which physical variables obey Langevin or related stochastic equations.
• Spin-based devices: Voltage-controlled spin-orbit-torque magnetic tunnel junctions (VC-SOT MTJs) where voltage pulses modulate magnetic anisotropy barriers and spin-orbit-torque currents programically bias network states, enabling Boltzmann machine operation with fJ-scale energy per operation (Lee et al., 2021).
• Quantum thermal machines: Autonomous few-qubit systems coupled to unbalanced temperature baths, realizing "thermodynamic neurons" that compute Boolean functions using engineered heat flows and modulated steady states of auxiliary reservoirs (Lipka-Bartosik et al., 2023).
• Hybrid classical–thermodynamic architectures: Conventional digital CPUs interface with thermodynamic coprocessors (e.g., stochastic processing units) to offload matrix computations or generative sampling tasks (Melanson et al., 2023, Bartosik et al., 2024).

Each architecture exploits natural stochastic dynamics—via noise-injected analog hardware, energy landscape manipulation, or stochastic resonance—to perform computation efficiently, often as the system approaches or samples from a thermodynamically defined equilibrium or steady state (Melanson et al., 2023, Whitelam et al., 2024).

3. Algorithms and Optimization in Thermodynamic Computing

Thermodynamic computing supports a variety of computational primitives and algorithms:

• Logic Gate Implementation: By programming couplings between "visible" (information-bearing) variables and ancillary "control" units, logic gates such as erasure, XOR, NOR, and majority functions are realized via static (time-independent) potentials. Genetic algorithms can be used to train the couplings for high-fidelity logic with optimized, program-dependent energy dissipation. Dissipation can even be redirected from the logical bits to the control unit, enabling programmable heat management such that the ratio of control to bit heat is O(1), a dramatic improvement over CMOS or feedback-trap implementations where it is O(10⁴–10⁸) (Whitelam, 13 Feb 2026).
• Matrix Algebra and Linear System Solving: Langevin or Ornstein–Uhlenbeck dynamics in quadratic energy landscapes allow matrix inversion, exponentials, and determinant estimation by reading out equilibrium or time-correlated covariance statistics. For instance, the covariance $\langle x x^T \rangle$ of the process is proportional to $A^{-1}$, accelerating tasks such as support vector machine training and quadratic programming. Thermodynamic algorithms deliver asymptotic quadratic ($O(d^2)$) speedup for full $d \times d$ matrix operations over classical digital routines ($\tilde{\gamma}$0), leveraging "thermodynamic parallelism"—the simultaneous statistical evolution of all $\tilde{\gamma}$1 correlation channels under noise (Bartosik et al., 2024, Duffield et al., 2023).
• Universal Function Approximation: Nonlinear function computation (beyond Boolean logic) is enabled via networks of thermodynamic neurons—oscillators with quartic or higher potentials—trained by genetic algorithms to approximate arbitrary continuous maps at specified observation times, either in equilibrium or out-of-equilibrium regimes (Whitelam et al., 2024). This enables general-purpose computation akin to deep neural networks, driven by thermal noise.
• Generative Modeling and Diffusion: Generative thermodynamic computing utilizes overdamped Langevin dynamics and parameterized potentials to synthesize structured data from thermal noise. Training the potential landscape to maximize the reverse likelihood of noising trajectories ensures minimal heat emission per sample, paralleling classical diffusion models but with physically interpretable energy costs (Whitelam, 18 Jun 2025).

4. Thermodynamic Resource Tradeoffs and Error Management

A central concern is the explicit tradeoff among accuracy, energy, and time:

• Energy-Time-Accuracy Bound: The energy-delay-deficiency product (EDDP) bounds the irrecoverable work, total time, and fraction of computation spent outside sampling (deficiency). For overdamped Langevin systems,

$\tilde{\gamma}$2

where $\tilde{\gamma}$3 is the Wasserstein distance between distributions, setting a universal lower bound on thermodynamic resource consumption (Rolandi et al., 7 Jan 2026).

• Control Protocol Design: Geodesic (Wasserstein-optimal) protocols minimize dissipation but require prior knowledge of the output distribution. Quasi-optimal continuous ramps and bang-bang schemes can achieve practical near-optimal tradeoffs without such knowledge.
• Error Mitigation: Analog hardware introduces finite precision, leading to unavoidable rounding errors in, e.g., programmable covariance matrices. The Thermies protocol reduces the dominant error from $\tilde{\gamma}$4 to $\tilde{\gamma}$5 via ensemble sampling of near-neighbor physical distributions, with demonstrated 20% error reduction for 8x8 matrix inversion in real devices and verified scalability beyond $\tilde{\gamma}$6 (Aifer et al., 2024).

5. Thermodynamic Limits, Circuit Complexity, and Logical Universality

• Thermodynamic Cost of Turing Machines: For any computable realization of a universal TM, the thermodynamic complexity (minimum heat to generate any desired output) is bounded by a constant $\tilde{\gamma}$7, in sharp contrast to unbounded Kolmogorov complexity. However, the average heat for entropy-minimizing input distributions diverges, indicating a tradeoff between reversibility, physical heat, and algorithmic complexity (Kolchinsky et al., 2019).
• Circuit Topology and Modular Dissipation: The dissipation in modular digital circuits is governed not just by local Landauer bounds but also by a "Landauer loss"—additional entropy production due to isolation and reinitialization of subcircuits, quantifiable via mutual information loss between gates and their environment. Optimization of circuit topology, alignment of gate priors, and reversible gate protocols are required for low-dissipation design (Wolpert et al., 2018).
• Minimality via Language Theory: Among all DFAs recognizing the same language, the minimal DFA is always the least dissipative for any decomposition and any bath dynamics, due to monotonicity of Shannon entropy under coarsening of state partitions. This provides a fundamental guide for thermodynamic computing architectures to prioritize state-space and logic minimization (Kardeş et al., 2022).

6. Outlook: Applications and Open Directions

TC offers unique capabilities and open questions:

• AI and Neuromorphic Applications: Thermodynamic co-processors are being deployed for probabilistic inference, generative modeling, and uncertainty quantification, promising significant scaling advantages and ultra-low power operation for large Gaussian or diffusion-based models (Melanson et al., 2023).
• Optimization and Simulation: Hybrid digital–thermodynamic algorithms for quadratic programming, portfolio optimization, and circuit simulation achieve polynomial speedup and high accuracy by marrying digital pre/post-processing with analog nonequilibrium relaxation (Bartosik et al., 2024).
• Stochastic Reservoir Computing: Harnessing thermodynamic fluctuations and timing instabilities in unconventional substrates (e.g., voltage-stressed ASICs) introduces new forms of efficient physical reservoir computing that may scale logarithmically in energy with state size (Lafuente et al., 5 Jan 2026).
• Programmable Heat Management: TC architectures enable explicit program-level targeting not just of computational accuracy or speed but the spatial distribution of heat emission, setting a foundation for computational tasks with embedded thermal management as a first-class design objective (Whitelam, 13 Feb 2026).
• Open Challenges: Central research questions include development of non-equilibrium theories for adaptation and learning, scalable materials and device engineering for robust stochasticity, rigorous complexity-theoretic classifications embedding thermodynamic resources, and integration of machine learning frameworks for programmable TC devices (Conte et al., 2019).

Thermodynamic Computing thereby establishes a physically-founded, resource-aware extension of computation in which dissipation, adaptation, and noise are computational resources rather than limitations, enabling both new fundamental theoretical developments and practical advances in ultra-efficient information processing.