Thermodynamic computing
 | This article has multiple issues. Please help improve it or discuss these issues on the talk page. (Learn how and when to remove these messages)
|
Thermodynamic computing refers to a type of computing which precursors and precedents have existed for decades for specialized uses in stochastic computing.[1] Then, in the early-2020s, the field began undergoing accelerating development for broader uses, driven by the computing needs of artificial intelligence, especially via pioneering work by the computing companies Extropicand Normal Computing.[2] [3]
History
[edit]Background
[edit]Stochastic computing was investigated as early as the 1960s and 1970s, when engineers proposed circuits that performed stochastic sampling rather than fixed Boolean logic. Boltzmann machines based on statistical mechanics and energy-based neural networks provided the theoretical foundation for using physical energy landscapes to represent probability distributions. This was also developed further in machine-learning research on diffusion and generative models.
In the 2000s and 2010s, developments in quantum annealing, notably D-Wave Systems computers and memristive systems, further demonstrated how physical systems could relax toward low-energy states corresponding to computational solutions. Extropic's approach represents a continuation of this tradition, replacing fully digital logic with thermodynamic sampling units (TSUs) designed to exploit controlled fluctuations for energy-efficient inference.
Computing structure
[edit]TSUs operate differently than conventional CPUs; instead of processing a series of programmable deterministic computations, TSUs produce samples from a programmable distribution. [3]
Thermodynamic computing samples from complex probability distributions, omitting matrix multiplication TSUs sample from energy-based models (EBM), a type of machine learning model that directly define the shape of a probability distribution via an energy function. This distinguishes them from conventional AI algorithms that are based on sampling from complex probability distribution; current AI systems generally produce a vector of probabilities, and then derive a sample from that.
The inputs to a TSU are parameters that specify the energy function of an EBM, and the outputs of a TSU are samples from the defined EBM. To use a TSU for machine learning, the parameters of the energy function are adjusted so that the EBM on the actual TSU will constitute a reliable model of real-world conditions. [3]
Hardware development
[edit]As of 2026[update], at least two companies are pursuing thermodynamic computing hardware and software, both founded in the United States in 2022: Extropic[4][5] and Normal Computing.[6][7] [3]
One article notes:
Normal Computing has announced its successful tape-out of the world's first thermodynamic computing chip, called CN101. Designed for AI/HPC data centers, the ASIC is a step away from traditional silicon computation methods that uses thermodynamics (and other physics principles) to reach computational efficiency that traditional chips can't match.
Thermodynamic chips are a world apart from traditional computing — closer in practice to the realms of quantum and probabilistic computing. Where noise is the enemy of standard electronics, thermodynamic and probabilistic chips actively use noise to solve problems.
“We’re focusing on algorithms that are able to leverage noise, stochasticity, and nondeterminism,” said Zachary Belateche, silicon engineering lead at Normal Computing, in a recent interview with IEEE Spectrum. “That algorithm space turns out to be huge, everything from scientific computing to AI to linear algebra."[8]
See also
[edit]References
[edit]- ^ Argent-Katwala, Ashok; Bradley, Jeremy T. (2006) [21-22 June 2006]. "Functional Performance Specification with Stochastic Probes". Written at Budapest, Hungary. In Horváth, András; Telek, Miklós (eds.). Formal Methods and Stochastic Models for Performance Evaluation (EPEW 2006). Lecture Notes in Computer Science. Vol. 4054. Berlin, Heidelberg: Springer Science+Business Media. pp. 31–46. doi:10.1007/11777830_3. ISBN 978-3-540-35362-1. Retrieved 3 March 2026. free PDF version
- ^ "Thermodynamic Computing: From Zero to One". Extropic Corp. 29 October 2025. Retrieved 3 March 2026.
- ^ a b c d Melanson D, Abu Khater M, Aifer M, Donatella K, Hunter Gordon M, Ahle T, Crooks G, Martinez AJ, Sbahi F, Coles PJ. Thermodynamic computing system for AI applications. Nat Commun. 2025 Apr 22;16(1):3757. doi: 10.1038/s41467-025-59011-x. PMID: 40263283; PMCID: PMC12015238.
- ^ Knight, Will (29 October 2025). "Extropic Aims to Disrupt the Data Center Bonanza". Wired. San Francisco, California, United States. Retrieved 3 March 2026.
- ^ Mohammed, Viquar Younus (26 November 2025). "When Silicon Meets Chaos: Inside Extropic's Thermodynamic Bet on the Future of AI". Medium. United States. Retrieved 3 March 2026.
- ^ Staff (2022–2026). "Normal Computing: AI for our most pressing crises in silicon". Normal Computing Corp. New York, New York, United States. Retrieved 3 March 2026.
- ^ Melanson, Denis; Khater, Mohammad Abu; Aifer, Maxwell; Donatella, Kaelan; Gordon, Max Hunter; Ahle, Thomas; Crooks, Gavin; Martinez, Antonio J.; Sbahi, Faris; Coles, Patrick J. (22 April 2025). "Thermodynamic computing system for AI applications". Nature Communications. 16 (3757). United States. doi:10.1038/s41467-025-59011-x. Retrieved 3 March 2026.
- ^ Grimm, Sunny (13 August 2025). "World's first 'thermodynamic computing chip' reaches tape out – Normal Computing's physics-based ASIC changes lanes to train more AI". Tom's Hardware. Future US. Retrieved 4 March 2026.
