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 sample

Its most important lesson is not the promotional 10,000x number by itself. It is the decomposition of evidence:

  1. a random-bit circuit primitive is measured;
  2. DTM workloads are simulated;
  3. a future full accelerator is physically modeled;
  4. 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

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.