Denoising Thermodynamic Models
Summary
Denoising Thermodynamic Models (DTMs) are generative models that compose several conditional energy-based models as reverse-denoising steps. The corresponding Denoising Thermodynamic Computer Architecture (DTCA) maps those steps to sparse, locally connected Boltzmann-machine hardware driven by all-transistor stochastic sampling cells.
Role In The Wiki
DTM is the current hardware-algorithm co-design anchor for combining diffusion-style gradual denoising with energy-based models that are intentionally easy to sample on probabilistic hardware.
simple noise
-> conditional hardware EBM chain
-> generated sampleIts most important lesson is not the promotional 10,000x number by itself. It is the decomposition of evidence:
- a random-bit circuit primitive is measured;
- DTM workloads are simulated;
- a future full accelerator is physically modeled;
- the energy advantage is projected on binarized Fashion-MNIST.
Those four layers must remain separate until a complete DTCA device is fabricated and benchmarked end to end.
Official Artifacts
- Journal article: https://www.nature.com/articles/s44335-026-00075-3
- DOI: https://doi.org/10.1038/s44335-026-00075-3
- Preprint: https://arxiv.org/abs/2510.23972v2
- Official reproduction code: https://github.com/pschilliOrange/dtm-replication
- Official THRML library: https://github.com/extropic-ai/thrml
- Official X announcement: https://x.com/extropic/status/2072798208133128255
Evidence
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in denoising-thermodynamic-models-2026 rather than duplicating verdict rows here.
At the entity level, DTMs are an adjacent candidate substrate for probabilistic time-series generation and energy-efficient sampling. No current experiment establishes multivariate numeric fidelity, irregular-time support, calibrated future modes, context conditioning, or action-conditioned rollout.