An efficient probabilistic hardware architecture for diffusion-like models

Source

• Raw Markdown: paper_denoising-thermodynamic-models-2026.md
• Version-of-record PDF: paper_denoising-thermodynamic-models-2026.pdf
• Journal article: npj Unconventional Computing 3, 30 (2026)
• DOI: https://doi.org/10.1038/s44335-026-00075-3
• Preprint: arXiv:2510.23972v2
• Official X announcement: https://x.com/extropic/status/2072798208133128255
• Official reproduction code: https://github.com/pschilliOrange/dtm-replication
• Official THRML simulation library: https://github.com/extropic-ai/thrml
• Supplementary information: Nature supplementary PDF

The authenticated X root, Crossref record, code snapshots, and artifact audit are preserved under papers/denoising-thermodynamic-models-2026/ as raw provenance rather than linked as public wiki artifacts.

Status And Credibility

This is a peer-reviewed, open-access Nature Portfolio journal article published on 2026-07-02 after acceptance on 2026-05-28. The team combines Extropic researchers with MIT professor Isaac L. Chuang. The paper provides a version of record, measured random-bit circuit behavior, explicit system-model assumptions, public benchmark data, public reproduction code, and an open-source simulator.

The credibility boundary is equally important. Extropic funded the work, some authors were part of the company, the company intends to commercialize the technology, and it reports a provisional patent application covering the manuscript. The strongest system claim is a modeled projection for a future accelerator, not an independently reproduced end-to-end hardware measurement.

Core Claim

A monolithic energy-based model (MEBM) faces a mixing–expressivity tradeoff: making one energy landscape expressive enough for multi-modal data also makes local sampling between modes slow. A Denoising Thermodynamic Model (DTM) instead spreads generation across a chain of simpler conditional EBMs. Each reverse step solves a narrower denoising problem whose energy landscape can remain easier to sample.

For a reverse step, the paper uses

$$P_{\theta }(x^{t-1}\mid x^t)\propto \sum _{z^{t-1}}\exp \left[-\left(E_{t-1}^f(x^{t-1},x^t)+E_{t-1}^{\theta }(x^{t-1},z^{t-1};\theta )\right)\right].$$

$E_{t-1}^f$ keeps the proposed clean state near the noisy input, while $E_{t-1}^{\theta }$ supplies learned local structure through latent variables. The Denoising Thermodynamic Computer Architecture (DTCA) maps these sparse, locally connected Boltzmann-machine steps onto arrays of all-transistor stochastic sampling cells.

flowchart LR
  N[Simple noise x^T] --> E3[Hardware EBM step T]
  E3 --> E2[Hardware EBM step T-1]
  E2 --> E1[Hardware EBM step 1]
  E1 --> X[Generated sample x^0]
  R[Measured random-bit circuit] -. primitive and energy model .-> E3
  R -.-> E2
  R -.-> E1

Training And Architecture Notes

• A DTM composes multiple EBMs rather than asking one MEBM to represent the full data distribution.
• Sparse local connectivity makes blocked Gibbs sampling physically implementable, but also constrains which correlations each hardware EBM can represent.
• The Adaptive Correlation Penalty (ACP) monitors sampler autocorrelation and adjusts a total-correlation penalty so each denoising layer remains tractable to sample during training.
• The paper’s strongest Fashion-MNIST models chain two to eight EBM steps. Increasing depth, connectivity, allowed mixing time, or some model dimensions can help, but the gains do not scale monotonically without jointly increasing connectivity and sampling effort.
• The proposed longer-term path is hybrid thermodynamic-deterministic machine learning: deterministic neural networks encode rich observations into a hardware-friendly latent space, while DTMs perform efficient probabilistic generation in that latent space.

What The Paper Directly Demonstrates

Modeling and sampling

On binarized Fashion-MNIST, simulated DTMs train more stably and reach better FID at much shorter mixing requirements than the paper’s monolithic Boltzmann-machine baseline. ACP prevents the late training instability seen in an unregularized DTM by keeping layer autocorrelation low.

Hardware primitive

The authors measure an all-transistor random-bit circuit with an approximately sigmoidal programmable output probability and roughly exponential autocorrelation decay. This validates a key sampling-cell primitive. The measured chip is explicitly a circuit test vehicle: it does not implement an Ising machine or the full DTCA.

System energy projection

The headline ~10,000x result combines the measured random-bit behavior with circuit/process simulations, a physical energy model, and simulated DTM workloads. It estimates that a future DTCA device could match selected GPU generative baselines at a given Fashion-MNIST FID while using roughly four orders of magnitude less inference energy per sample.

This is not a measured GPU-versus-DTCA device comparison. The DTM experiments run as simulations on conventional hardware, and the energy model omits or simplifies effects that a real implementation must confront.

Preliminary hybrid model

A small CIFAR-10 experiment combines a learned deterministic image embedding with a DTM. At the reported matched quality, the hybrid uses a deterministic neural component with about one tenth the parameters of the compared GAN. The paper calls this a naive first attempt: the encoder and DTM are not jointly trained, and the result is far from frontier image-generation quality.

Social Claim Audit

The user-provided Extropic post says that hardware built on primitives validated in X0 “could unlock generative modelling at 10,000x lower energy than GPUs.” The cautious paper-supported reading is:

Scope	Exact status
Publication claim	Verified: version of record in npj Unconventional Computing, published 2026-07-02.
Primitive validation	Real laboratory measurements support the random-bit circuit primitive.
Full hardware system	Proposed and physically modeled; no working DTM/DTCA chip is evaluated.
10,000x energy number	Modeled inference estimate on binarized Fashion-MNIST at matched FID, not direct end-to-end hardware metrology.
General generative modeling	Research direction only; current experiments are small image benchmarks, and raw capability is far below state of the art.

Public Artifact Surface

• pschilliOrange/dtm-replication is named by the paper’s Code availability statement and was pinned at commit 7c22d19c218ab353a770f7c2b1504570b7cbe3ba during ingest.
• extropic-ai/thrml is the underlying Apache-2.0 JAX simulation library and was pinned at commit 9c4e6fbb800f5e5c627122e668ff1b158ef3782b during ingest.
• The reproduction repository covers MNIST/Fashion-MNIST training and figure-generation workflows, but its README says the grayscale graph manager differs algorithmically from the original research code. It is therefore a useful reproduction surface, not a bit-identical archive of every experiment.
• Fashion-MNIST and CIFAR-10 are public. No dataset payloads are stored in this repository.

Limitations And Gotchas

• The ~10,000x result is a system-model estimate grounded in one measured circuit primitive, not a measurement from a complete probabilistic accelerator.
• The energy model simplifies external communication, weight loading, analog nonlinearity, drift, calibration, process variation, quantization, and other implementation effects.
• The authors use an optimistic FLOP/joule GPU model and evaluate a small binarized image benchmark; this is not a matched wall-clock, fabrication-node, compiler, batch-size, or production-serving comparison.
• Raw modeling quality remains far behind modern image generators. The paper explicitly says the proposed system is not a finished product and that capability scaling is unresolved.
• Sparse, local, hardware-friendly EBM topology may not match the correlation structure of complex data. Simply adding nodes can make mixing worse unless connectivity and sampling effort also scale.
• Binarization is not a general interface for rich observations. The preliminary hybrid deterministic encoder is not jointly trained with the DTM.
• The paper does not evaluate numeric time series, multivariate constraints, irregular sampling, event streams, context conditioning, actions, control inputs, interventions, or counterfactual prediction.
• Corporate funding, commercialization plans, and the patent disclosure increase the need for independent replication of the system-level estimate.

Foundation TSFM Relevance

Agenda slot	Verdict	Evidence	Missing pieces
Multi-modal future distributions	adjacent	DTM composes conditional latent-variable EBMs to avoid one rugged monolithic energy landscape.	No conditional numeric future trajectories, calibrated regime probabilities, rare-tail tests, or action-sensitive probability shifts.
Generation and editing	adjacent	Demonstrates diffusion-like image generation through a chain of hardware-oriented EBMs.	Needs multivariate time-series generation, dense numeric fidelity, irregular time, constraints, and editing tests.
Scaling substrate	adjacent	Provides a hardware-algorithm co-design path and a large modeled energy advantage for one small benchmark.	Needs fabricated end-to-end hardware, matched system benchmarks, richer data, wall-clock/throughput measurements, and scaling curves.
Benchmarks	warning	Separates measured circuit behavior, simulated models, and projected system energy, showing why these evidence layers must not be collapsed.	Future reports need end-to-end device measurements and matched production baselines.
Control and counterfactuals	insufficient evidence	The paper mentions robotic path planning only as EBM background and markets future world-model uses, but evaluates no action channel.	Add explicit actions/control inputs, intervention-conditioned rollout, calibrated outcomes, and planner utility.

Links Into The Wiki

• Denoising Thermodynamic Models
• Energy-Based Models
• Time-Series Generation
• Time-Series Scaling And Efficiency
• Foundation Time-Series Model Research Agenda
• Contradictions And Open Tensions

Open Questions

• Does the measured random-bit primitive retain its energy and sampling advantages after a complete DTCA chip adds routing, memory, weight programming, calibration, and host communication?
• Can DTMs scale beyond local sparse Boltzmann-machine topologies without recreating the mixing problem or letting communication dominate energy?
• Can deterministic encoders and DTMs be trained jointly while preserving calibrated probability mass and dense numeric detail?
• What is the fairest hardware comparison protocol: matched FID, wall-clock, throughput, fabrication node, batch size, latency, energy-delay product, or full training-plus-serving cost?
• Can a DTM generate calibrated multivariate time-series futures under context and explicit actions/control inputs, rather than only unconditional images?