Extropic

• Thermodynamic Sampling Units designed specifically for probabilistic AI workloads
• Z1 thermodynamic accelerator architecture with more than 269,000 pbits on a single chip
• Z1 Stick accelerator planned in an M.2 form factor for low-power edge applications
• Z1 Card PCIe accelerator planned for servers and workstations with more than four million pbits
• Torx open-source framework for general stochastic differentiable programming
• THRML framework for thermodynamic hypergraphical models and probabilistic graphical model simulations
• GPU-based simulation API for experimenting with thermodynamic programs before dedicated hardware availability
• Probabilistic primitives including pbits, pdit, pmode, and pmog for stochastic computation
• Support for AI research areas including diffusion models, simulations, world models, optimization, and probabilistic inference
• Thermodynamic computing stack combining silicon, systems, compilers, programming frameworks, and accelerator hardware

Extropic Overview

Extropic is a computing company developing thermodynamic hardware and software for probabilistic artificial intelligence workloads. Its approach replaces conventional deterministic computing primitives with probabilistic hardware designed to sample directly from probability distributions. The company describes its Thermodynamic Sampling Units, or TSUs, as a new type of computing hardware designed specifically for generative AI and other workloads that naturally involve probability and sampling.

Extropic publicly emerged from stealth in March 2024 after developing its thermodynamic computing technology and an initial superconducting prototype. The company later shifted significant development toward semiconductor-based thermodynamic computing that can operate at more practical temperatures and manufacturing processes. Its current technology stack spans silicon hardware, systems, compilers, stochastic programming frameworks, and simulation tools.

Thermodynamic Computing

Extropic’s core technology is based on thermodynamic computing, which uses physical sources of randomness to perform probabilistic computation. Instead of treating randomness as something that must be generated artificially by conventional digital processors, its hardware is designed to use naturally occurring thermodynamic fluctuations in electronic components to produce programmable probability distributions.

The company’s Thermodynamic Sampling Units are designed to sample directly from probability distributions. Extropic’s technical material connects this approach with generative AI because model inference and other probabilistic algorithms require sampling as part of their operation. The architecture is therefore intended to perform probabilistic operations closer to their physical basis rather than relying entirely on conventional deterministic digital computation.

Z1 Thermodynamic Accelerator

Z1 is Extropic’s first production-scale thermodynamic chip. The company reports that the chip contains 269,568 probabilistic bits, or pbits, arranged in a sparse graph and operates at a sampling rate above 50 MHz while consuming less than one watt. The chip is designed around energy-based models and performs chromatic Gibbs sampling directly in hardware.

Extropic has announced plans to develop Z1-based accelerator products in multiple form factors. The Z1 Stick is planned as an M.2 device containing more than half a million pbits for low-power edge applications such as robotics and augmented reality. The Z1 Card is planned as a PCIe accelerator containing more than four million pbits for servers and workstations.

Probabilistic Hardware

Extropic’s hardware architecture uses probabilistic primitives rather than conventional deterministic bits as the basic computational building blocks. Its current terminology includes pbits, pdit, pmode, and pmog, which represent different forms of programmable probabilistic behavior.

The company’s earlier X0 silicon prototype demonstrated that probabilistic circuits could be fabricated using standard semiconductor processes. Extropic subsequently developed the XTR-0 experimental platform around this technology, providing a physical system for researchers and early adopters to experiment with thermodynamic computing.

Torx Stochastic Programming

Torx is Extropic’s open-source framework for general stochastic differentiable programming. The framework is designed to provide a higher-level programming model for stochastic systems and is hardware-agnostic, allowing programs to be developed for stochastic processing units and simulated environments.

Torx represents stochastic programs as compositions of parameterized stochastic operations. This provides a programming approach for probabilistic systems that is conceptually similar to how conventional machine learning frameworks allow developers to construct differentiable programs, while targeting stochastic rather than purely deterministic computation.

THRML

THRML is Extropic’s open-source library for thermodynamic hypergraphical models. It uses JAX to accelerate simulations of probabilistic graphical models, including energy-based models intended for implementation on Extropic’s thermodynamic hardware.

Researchers can use THRML to experiment with probabilistic models, simulate potential thermodynamic sampling architectures, and explore different hardware configurations. Extropic also describes applications of the framework for training energy-based models and researching thermodynamic AI algorithms.

Simulation and API Access

Extropic provides an early-access API for simulating thermodynamic programs on high-performance GPU infrastructure. The simulator is intended to allow developers and researchers to experiment with Torx and THRML programs while the company’s larger-scale hardware is still being developed.

The simulation environment provides a way to investigate thermodynamic algorithms without requiring physical Z1 hardware. Extropic’s roadmap indicates that its software and simulation infrastructure will continue expanding toward more advanced accelerators, research tools, and eventual access to Z1-based computing clusters.

AI Applications

Extropic identifies several potential application areas for its thermodynamic computing platform. These include generative AI, diffusion models, scientific simulations, optimization, probabilistic inference, financial modeling, robotics, autonomy, and world models.

The company is particularly focused on workloads where probabilistic computation is fundamental. Its published application areas include text, image, and video diffusion models; nuclear, chemical, and molecular simulations; weather and robotics world models; discrete and continuous optimization; and Bayesian filtering and financial modeling.

Energy Efficiency Focus

Energy efficiency is a central objective of Extropic’s architecture. The company argues that conventional AI accelerators perform many probabilistic workloads indirectly through deterministic operations and data movement. Its thermodynamic approach instead attempts to perform sampling directly through physical probabilistic processes.

Extropic reports that its thermodynamic circuits can generate programmable random samples with substantially lower energy requirements than comparable deterministic approaches. The company has also reported potential large efficiency improvements for suitable stochastic workloads, although these figures are dependent on the workload, implementation, and comparison methodology.

Hardware Roadmap

Extropic’s current roadmap extends from experimental X0 and XTR-0 systems toward Z1 production-scale chips and larger thermodynamic computing systems. The company has stated that Z1 Stick and Z1 Card products are planned for early access in 2027.

The planned Z1 ecosystem includes edge-oriented accelerator sticks, PCIe cards for conventional servers and workstations, and eventually large-scale thermodynamic computing clusters. Extropic has described a longer-term goal of building systems containing one billion or more pbits for researching thermodynamic AI algorithms at substantially larger scales.

Research and Open Source

Extropic makes several of its software tools available as open-source projects and publishes technical research describing its hardware and programming approach. Its software ecosystem is intended to encourage researchers and developers to investigate stochastic programming, energy-based models, probabilistic computing, and thermodynamic AI before the company’s larger hardware systems become broadly available.

The company also operates an open research program and invites researchers to explore applications of probabilistic and thermodynamic computing. Its published application areas span AI generation, scientific modeling, optimization, simulation, and probabilistic reasoning.

Pricing and Availability

Extropic does not publish a standard commercial subscription price for its thermodynamic computing platform. Its open-source software frameworks are available for experimentation, while the company’s physical Z1 accelerator products remain under development. The company has announced early-access plans for Z1 Stick and Z1 Card hardware in 2027, but public starting prices have not been disclosed.

Frequently Asked Questions

What is Extropic?

Extropic is a computing company developing thermodynamic hardware and software for probabilistic AI workloads.

What is thermodynamic computing?

Thermodynamic computing uses physical thermodynamic fluctuations to perform probabilistic computations and generate programmable samples from probability distributions.

What is a Thermodynamic Sampling Unit?

A Thermodynamic Sampling Unit is Extropic’s proposed computing architecture designed to sample directly from programmable probability distributions for probabilistic workloads.

What is the Z1 chip?

Z1 is Extropic’s first production-scale thermodynamic chip, containing 269,568 pbits and designed for energy-based probabilistic computation.

What is Torx?

Torx is Extropic’s open-source framework for general stochastic differentiable programming and is designed to work across different stochastic processing environments.

What is THRML?

THRML is an open-source library for thermodynamic hypergraphical models that supports probabilistic graphical model simulation and energy-based model research.

Is Extropic hardware available for purchase?

Extropic’s larger Z1 hardware is not generally available for purchase, with Z1 Stick and Z1 Card early access planned for 2027.

Does Extropic have public pricing?

No, Extropic has not publicly disclosed starting prices for its planned commercial thermodynamic hardware or future production-scale systems.

Share Your Experience: