This project runs its architecture on GPUs. The question on this page is whether the result would change on other hardware. If the architecture’s measured properties survive a move to a different physical substrate, such as spiking neuromorphic chips or thermodynamic p-bit samplers, that is evidence for substrate independence. If they do not survive, that is evidence against it. Either result is worth having.
This page is the hub of the thermodynamic computing section. The console below is a software emulation. The posts further down cover the hardware, the physics and the energy budgets around it. The hardware kits page lists parts to build a p-bit sampler on a desk.
The success criterion is narrow on purpose. The test is whether the properties the project has already verified survive the change. Those properties are the vector representations that carry the stimulus, and the chain from the observation map to the tectum content. One limit applies to them. They were measured on a closed set of 72 stimuli, and the test stimuli also occurred in training. So they show that a known stimulus is told apart from the others. They do not yet show that a new stimulus is read. It is not a claim that a new substrate improves any emergence metric. None of the project’s 21 instrument entries is trusted yet, so no such claim could be checked.
The theory this architecture follows does not settle the question in advance. Todd Feinberg and Jon Mallatt, whose six neurobiological features shape the design, argue in a 2026 commentary in Behavioral and Brain Sciences that the physical substrate of a complex system is critical for sentience to emerge. On their view, a change of substrate can change more than speed and cost. The project uses their features as an architectural blueprint and does not claim their support for substrate independence. A positive result here would show that some measured properties survive the move. It would not show that the new substrate feels anything.
This visualization runs a software emulation in your browser. Physical thermodynamic hardware is not integrated. Full statement.
96 oscillators on a circle, the case the workspace binding runs. Below K of about 1.6 they drift, and above it they lock. An orange dot has moved less than a quarter turn against the mean field over the last 90 frames, so it keeps step with it. A pale dot keeps slipping past it.
96 oscillators on a sphere. The binding code supports this, and the running architecture does not use it. There is no threshold. Drag to turn the sphere.
On the left, stochastic sampling while the gain rises. On the right, descent with no noise, which flips a spin only when the flip lowers the energy. Both start from the same lattice and the same bonds. In the spin glass every bond has a random sign, so no arrangement satisfies them all, and descent stops in the first local minimum it reaches. Cool slowly and the sampler usually ends lower. Cool fast and it behaves more like descent.
Ising lattice. Each square is a p-bit, a spin that is either up (orange) or down. A spin looks at its four neighbors and the field, then goes up with probability (1 + tanh(h / T)) / 2. That is the update rule for p-bits from Camsari and colleagues. Near the Onsager temperature, 2.269, the lattice forms domains at every scale. Below it, one direction wins. Neighboring spins never update at the same time. The lattice is split into a checkerboard, and the two colors take turns. Updating everything at once is faster to write and samples the wrong distribution.
Kuramoto ring and sphere. Each dot is an oscillator with its own natural frequency. Coupling pulls them together, and the length of the white line is the order parameter r. On the circle (D = 2), the oscillators drift apart below a critical coupling of about 1.6 and lock above it. The project’s workspace binding runs exactly this case, with its oscillators on a circle, following the artificial Kuramoto oscillatory neurons of Miyato and colleagues.
The sphere view (D = 3) shows the same rule with each oscillator as a point on a sphere, turned by its own rotation. The binding code supports this, but the running architecture does not use it. The sphere behaves differently. It has no threshold, and order appears as soon as the coupling is switched on, as Chandra, Girvan and Ott showed for odd dimensions. Drag the sphere to turn it.
Sampling vs descent. Both lattices start from the same random state. Descent without noise freezes quickly with domain walls it cannot remove. Sampling with a slowly rising gain usually removes more of them and reaches a lower energy. Thermodynamic hardware is built around this property. Its noise lets a system leave a poor local minimum.
The ferromagnet is the easy case, because its best arrangement is known: every spin the same way, energy -2 per spin. The second problem is harder. In the two dimensional spin glass each bond has a random sign, so no arrangement can satisfy every bond, and the energy landscape has many local minima. Exact computations on large lattices put its lowest energy at about -1.40 per spin (De Simone and colleagues 1996 give -1.4015), and the plot draws that line. A single 48 by 48 sample scatters around it. The cooling rate and the seed are yours to change, so you can check whether the sampler’s advantage holds for a given start, and how it shrinks as the cooling gets faster.
Biological neural dynamics are continuous, asynchronous and parallel. A GPU simulation cuts time into steps of fixed size, so the cost grows with the number of steps times the number of units. Each belief update in a trained model is a forward pass, a backward pass and a weight update. A data center GPU draws between 250 and 700 W. The human brain runs on about 20 W.
The brain is not free to run. It pays in metabolic energy for every spike. The difference is in the cost structure. In tissue the dynamics are the physics, and in a GPU the physics has to be computed.
Karl Friston’s free energy principle describes living systems as minimizing variational free energy, a quantity that bounds surprise. It is tempting to say that thermodynamic hardware does this physically. That is not correct as stated, and the reason is below.
A p-bit chip samples from a Boltzmann distribution of whatever energy function is programmed into it. That is a fact about the hardware. It becomes inference only after an inference problem is compiled onto the chip’s couplings. Whether this project’s generative model can be compiled that way, and at what approximation cost, is an open question. It is a hypothesis this project can test. It is not a property the hardware already has.
| Level | Status | Technologies | Evidence |
|---|---|---|---|
| 1. Production | Deployed and measured | GPU tensor emulation, surrogate gradients, Brian2 comparison | Behavioral task metrics and perception probes in the research repository; instrument statuses in the inventory (0 of 21 trusted) |
| 2. Neuromorphic | Research systems, no data from this project | Intel Loihi 2, SpiNNaker 2, BrainScaleS-2, the NIR exchange format | Published third-party benchmarks |
| 3. Thermodynamic | Early hardware for early partners, open-source simulators | Extropic X0 and Z1 (Z1 early access planned for 2027), Normal Computing CN101 | Vendor figures; open-source THRML, Torx and thermox simulators |
The second level, neuromorphic hardware, is explained in a short film about spiking neurons and brain-like chips. Its Hala Point numbers are Intel’s figures for a research system.
Extropic states its Z1 chip carries 269,568 p-bits in eight cores, with 215,904 programmable couplings, a sparse graph of 16 neighbors per p-bit, sampling above 50 MHz, and under 1 W per die. Those are vendor figures. This project has not measured them. For the civilization scale framing of that efficiency term, see the Kardashev scale as an energy budget for growing consciousness.
The research code has an isolated package for this work,
models/thermodynamic/.
Phases 1 and 2 were merged on 2026-09-22. Phase 1 holds a p-bit emulator that updates one
color class of the coupling graph at a time, tested against exact Boltzmann statistics, and a
free energy relaxation solver tested against closed-form posteriors. Phase 2 adds a bridge
between continuous values and spike trains, a stand-in driver for tests, export of spiking
layers to the NIR exchange format, and an optional link to Extropic’s open-source THRML sampler,
checked against the same exact statistics. A guard test keeps the package out of every
production path, so nothing on this page can change a training run.
No result from the package is reported here as evidence for substrate independence. Since Phase 2 the research code has added an offline probe that sends recorded vectors through an emulated p-bit channel and reads them with the project’s own decoders. That work is in progress, and its documents are in the research repository. Physical hardware is not part of the plan until it is available.
Each tool below runs in software on a CPU or GPU. None of them needs thermodynamic hardware. Checked on 2026-10-27. The project’s research code calls THRML through an optional wrapper. A walkthrough of THRML, with a run checked against an exact answer, is in Extropic THRML, the open source thermodynamic sampler.
| Tool | From | What it does | Language | Licence | Latest release |
|---|---|---|---|---|---|
| THRML | Extropic | Block Gibbs sampling of discrete energy-based models, including Ising models, the scheme Z1 is stated to run | Python, JAX | Apache-2.0 | 0.1.4, August 2026 |
| Torx | Extropic | Stochastic circuits and factor graphs of kernels, the programs Thermalizers is meant to compile for Z1 | Python, JAX | Apache-2.0 | 0.0.3, October 2026 |
| thermox | Normal Computing | Exact simulation of Ornstein-Uhlenbeck processes and thermodynamic linear algebra | Python, JAX | Apache-2.0 | 0.0.6, October 2026 |
Extropic’s compiler, Thermalizers, was announced in Summer 2026 and was not yet public on the date above.
FPGA boards, true random number generators and an analog computer that can run the same p-bit, Ising and relaxation schemes on a desk are listed in the thermodynamic computing hardware guide.
The console is a software emulation that runs in your browser. It is not connected to any thermodynamic or neuromorphic chip, and it does not run the project’s architecture. The lattice is a textbook ferromagnet and the ring is the textbook Kuramoto model, chosen because their behavior is known exactly and you can check it by eye.
Nothing here is evidence that the architecture is substrate independent. That evidence can only come from running the verified parts of the architecture on another substrate and measuring them the same way. For the chip that has already flown, read about Intel’s Loihi in orbit. For the case for putting minds off planet, see Consciousness to Orbit. For a spiking network you can run now, open the Substrate Console.
Guillaume Verdon founded Extropic, which builds thermodynamic sampling hardware. Its efficiency figures are the company’s own claims. This project does not run on that hardware. Karl Friston wrote the free energy principle, the framework behind the section on thermodynamics and the free energy principle above.
For the Ising lattices and Kuramoto rings behind thermodynamic computing. Every order funds the research.
