Extropic unveils Thermodynamic Sampling Unit

Extropic’s Thermodynamic Sampling Unit, or TSU, is a bid to move part of AI computation away from deterministic arithmetic and toward direct physical sampling. The company says the hardware uses probabilistic bits, or p-bits, and thermal noise to generate samples from probability distributions in silicon rather than approximating those samples through conventional digital compute. Extropic has described the TSU publicly since October 2025, and a May 19 social-media post revived attention around the design and its use in Gibbs sampling for probabilistic models. The company’s public materials frame the hardware as a fit for sampling-heavy AI workloads, including diffusion-style systems. ### What exactly is a Thermodynamic Sampling Unit? Extropic defines a TSU as “a type of probabilistic computer” whose inputs specify a probability distribution and whose outputs are samples from that distribution. In the company’s explainer, the TSU is not presented as a faster GPU or CPU, but as hardware built around sampling as the primitive operation. The p-bit is the basic building block in Extropic’s description. (extropic.ai) Extropic says a p-bit is programmed with a probability and then outputs a 1 with that probability and a 0 otherwise, calling it a programmable analog hardware implementation of a Bernoulli distribution. Networks of those probabilistic circuits, linked through local communication, are what the company says let a TSU sample from more complex distributions useful for generative AI. ### Why does thermal noise matter here? Thermal noise is usually something chip designers try to suppress. Extropic’s pitch is the reverse: use those physical fluctuations as the randomness source. The company’s homepage says its hardware is “inherently probabilistic,” and its TSU explainer says the machine is designed to sample directly rather than simulate randomness through conventional digital methods. (extropic.ai) Communications of the ACM described thermodynamic computing in 2025 as an approach that harnesses stochastic thermodynamics and natural fluctuations instead of suppressing them. That broader framing matches Extropic’s public description, though the company’s specific implementation claims remain its own. ### Where does Gibbs sampling fit into the story? Gibbs sampling is one of the standard ways to draw samples from a complex probability distribution by updating variables from conditional distributions. (extropic.ai) Extropic’s May 19 post described the TSU as doing native Gibbs sampling for probabilistic graphical models, which fits the company’s longer-standing claim that its hardware is built around sampling rather than matrix multiplication. (cacm.acm.org) Extropic’s October 2025 explainer argues that modern generative AI was shaped by what GPUs are good at — matrix multiplication — and not because those operations are the natural form of every generative model. The company says TSUs could support algorithms that make fuller use of hardware-native sampling. ### How is this supposed to help diffusion models? Extropic’s October 28, 2025 paper proposed “an efficient probabilistic hardware architecture for diffusion-like models” and said an all-transistor probabilistic computer could implement denoising models at the hardware level. (extropic.ai) The paper’s system-level analysis said devices based on that architecture could reach performance parity with GPUs on a simple image benchmark while using about 10,000 times less energy. The paper used binarized Fashion-MNIST and compared denoising thermodynamic models with GPU baselines including variational autoencoders, GANs and DDPMs. That means the public energy claim is tied to a specific benchmark and system-level analysis, not to a broad demonstration across production-scale image or video models. ### Has Extropic shown production hardware yet? Extropic’s website says it has a prototype platform called XTR-0 and a software library called THRML for developing thermodynamic algorithms and simulating them on TSUs. (arxiv.org) The company’s public site also lists “Inside X0 and XTR-0” and “TSU 101” as its main technical explainers. As of May 19, the most detailed public technical sources remain Extropic’s own October 2025 materials and the company-authored arXiv paper. (arxiv.org) Those documents outline the architecture, the sampling model and the benchmarked diffusion-like use case; they do not amount to an independently verified demonstration of large-scale commercial deployment. (extropic.ai 1) (extropic.ai 2)