Thermodynamic Computing

last updated 2026-05-27
Probabilistic ComputingStochastic / Ising MachinesAnalog ComputingNeuromorphic ComputingThermodynam…

Verified 2026-05-27 against primary sources

Physics / mechanism

Thermodynamic computing exploits stochastic fluctuations in physical systems as the computational substrate. Rather than enforcing deterministic logic and dissipating energy to suppress thermal noise (as conventional CMOS does), thermodynamic chips engineer the noise itself to perform computation.

The mathematical model is Langevin dynamics — stochastic differential equations (SDEs) describing systems under combined deterministic and random forces. The hardware is built to be a direct physical realisation of these equations, so the chip’s natural evolution implements the sampling algorithm.

Key grounding: the fluctuation-dissipation theorem couples noise and dissipation, so a properly engineered noisy system can do useful Monte Carlo–style sampling work at energy levels far below what’s needed to fight noise as an error source.

Two distinct unit-cell families currently in silicon:

Both target the same workload class — sampling-heavy AI (diffusion models, Bayesian inference, energy-based models) plus stochastic scientific computing (molecular dynamics, materials simulation). What does not benefit: standard supervised deep learning, matrix multiplication for forward inference of conventional NNs.

Why now (as of May 2026)

  1. First silicon shipped. Normal Computing’s CN101 taped out June 2025 (2025 08 12 Normal Computing Cn101 Tape Out); Extropic’s X0 proof-of-concept Q1 2025; Extropic’s Z1 production chip targets early 2026. Both companies have moved from theory to silicon within the same window.
  2. Diffusion-model demand inflection. Image, video, and 3D generative workloads dominate consumer AI cost. Sampling these distributions on GPUs is the bottleneck thermodynamic chips claim to solve.
  3. Government endorsement. UK Government’s ARIA agency funded Normal Computing UK as 1 of 12 teams in the £100M Scaling Compute programme — explicit goal of 1000x AI hardware cost reduction (2024 10 28 Aria Scaling Compute Normal Computing). Suraj Bramhavar (ARIA programme director, ex-Sync Computing) personally co-authored Normal’s coalition paper (2025 07 14 Solving Compute Crisis Physics Based Asics).
  4. Strategic-LP money flowing. Samsung Catalyst Fund led Normal’s $50M strategic round March 2026 (2026 03 25 Normal Computing Samsung Catalyst 50M) — signals foundry interest. (Note: process node / foundry partner for CN101 remains undisclosed as of May 2026.)

Competitive landscape

ApproachCompaniesEnergy claimMaturityTarget workload
Thermodynamic continuous-stateNormal Computing~1000x vs GPU (press); “orders of magnitude” (their technical blog)CN101 silicon June 2025; CN201 2026Diffusion, Bayesian inference, matrix inversion
Thermodynamic / p-bit single-state (TSU)Extropic~10,000x vs GPU (system-level analysis)X0 proof Q1 2025; Z1 production early 2026Image / video generation, robotics control
CMOS p-bit (academic)Tohoku / Purdue / Aalto labs~1 fJ/op at room temp~1k p-bit research chipsCombinatorial optimisation, Bayesian sampling
Quantum annealingD-WaveμJ range with cryo overheadCommercial, >5000 qubitsCombinatorial optimisation
Classical Ising / annealerFujitsu DAU, Hitachi CMOS annealer, Toshiba SQBM+n/aCommercialQUBO / Ising-form optimisation
Neuromorphic / spikingIntel Loihi, BrainScaleS, SynSenseVariesResearch / pilotEvent-driven inference
GPU Monte Carlo (baseline)NVIDIA, AMDpJ/op, mature toolchainTRL 9General ML sampling

Companies using

Connected ideas

Sources

Frontier (open questions)

Concept page rewritten 2026-05-27 against primary sources. Previous content was Sonnet-auto-mapped 2026-05-04 and contained inaccuracies (described Normal’s units as “SPUs” — that’s the older PCB prototype; the silicon architecture is Carnot with “s-units”; said neither company had “disclosed volume silicon” — outdated since CN101 tape-out August 2025).

Related concepts

Frontier questions