Thermodynamic Compute Index
Primary-source index of thermodynamic, probabilistic, and analog AI compute
Thermodynamic and probabilistic AI compute draws samples from programmed distributions by using thermal noise, analog circuit dynamics, or probabilistic bits. This index records primary-source chips, boards, libraries, and papers, including Extropic X0 and Z1, Normal Computing's eight-cell stochastic processing unit, IBM's 14 nm analog-AI chip, and peer-reviewed p-bit machines, as retrieved on 19 October 2026.
60-Second Technical Overview & Benchmark Walkthrough
Verified Thermodynamic Compute Index Matrix (14 of 14 Tracked)
Backed by SQLite warehouse (warehouse/warehouse.db) · Last 30 days sensor verified
| Entity / Implementation | Organization | Status | Key Technical Metric | Governance / License | Verification |
|---|---|---|---|---|---|
| Extropic X0 Chip Chip Extropic's first silicon thermodynamic chip. Its pbits turn transistor thermal noise into programmable randomness. | Extropic | Prototype silicon | All-transistor pbits | Company disclosure | reported |
| XTR-0 Desktop System Platform Desktop probabilistic computer built around X0. Extropic says dozens were built and shipped to early adopters. | Extropic | Shipped to early adopters | Dozens of systems shipped | Company disclosure | reported |
| Extropic Z1 TSU Chip Extropic TSU figure: eight cores, 269,568 pbits, 215,904 couplings, above 50 MHz, and under 1 W. | Extropic | Taped out | 269,568 pbits | Company disclosure | reported |
| Z1 M.2 Compute Stick System Announced edge stick: two Z1 chips and over 500,000 pbits in an M.2 form factor, early access in 2027. | Extropic | 2027 early access | 500,000+ pbits | Company disclosure | reported |
| Z1 PCIe Accelerator System Announced server card: 16 Z1 chips and over 4,000,000 pbits in a PCIe form factor, early access in 2027. | Extropic | 2027 early access | 4,000,000+ pbits | Company disclosure | reported |
| THRML JAX Library Library Apache-2.0 JAX library for block Gibbs sampling of probabilistic graphical models. Requires Python 3.10+. | Extropic | Open source | Apache-2.0 | Apache-2.0 | verified |
| Torx Stochastic SDK Library JAX framework for parametrised stochastic circuits. The docs say to install extro-torx on Python 3.11+. | Extropic | Open source | Python 3.11+ | Open-source release | verified |
| Denoising Thermo Paper Paper Extropic preprint: simulated DTMs match GPU FID on binarized Fashion-MNIST at about 10,000× less energy. | Extropic | arXiv:2510.23972 | ~10,000× energy (sim.) | arXiv preprint | reported |
| Normal Computing SPU Board Normal Computing board with eight all-to-all RLC cells, used for Gaussian sampling and matrix inversion. | Normal Computing | Published 22 Apr 2025 | 8 all-to-all RLC cells | Nature Communications | verified |
| IBM Analog-AI Chip Chip IBM Research 14 nm analog-AI chip: 35 million PCM devices, 34 tiles, and up to 12.4 TOPS/W sustained. | IBM Research | Published 23 Aug 2023 | 12.4 TOPS/W | Nature | verified |
| UCSB Sparse p-bit DBM Machine UCSB FPGA p-computer: 4,264 p-bits reach 90% MNIST accuracy with about 30,000 parameters. | UC Santa Barbara | Published 17 Jun 2024 | 4,264 p-bits | Nature Electronics | verified |
| UCSB CMOS p-bit Chip Chip UCSB preprint of a 440-spin CMOS p-bit solver on a Chimera graph, occupying 0.44 square millimeters. | UC Santa Barbara | arXiv:2504.14070 | 440 spins, 0.44 mm² | arXiv preprint | reported |
| Northwestern VCMA ASIC Chip Northwestern 130 nm ASIC factors integers with 1,143 p-bits driven by VCMA magnetic tunnel junctions. | Northwestern University | Published 13 Aug 2025 | 1,143 p-bits, 130 nm | Nature Electronics | verified |
| Extropic Z1T Note Paper Extropic note maps transformer-like layers onto Z1 and states 2,135,904 hardwired coupling edges. | Extropic | Posted 4 Sep 2026 | 2,135,904 edges | Company technical post | reported |
Featured Head-to-Head Comparisons
Every pair of the 14 records (91 total pairs) has a static comparison page. These 6 are the highest-intent matches.
Frequently Asked Questions
What does the Thermodynamic Compute Index record?
It records thermodynamic, probabilistic, and analog AI compute from primary sources: Extropic chips and libraries, Normal Computing's stochastic processing unit, IBM Research's analog-AI chip, and peer-reviewed p-bit machines. Every row has a source URL retrieved on 2026-10-19.
What does Extropic state about the Z1 chip?
Extropic's Z1 figure caption lists eight cores, 269,568 pbits, 215,904 coupling parameters, a sampling rate above 50 MHz, and power under 1 W. The same post says each pbit connects to sixteen neighbors and that the chip samples an Ising model by chromatic Gibbs sampling.
What did Normal Computing demonstrate in hardware?
In Nature Communications on 22 April 2025, Normal Computing reported a stochastic processing unit with eight RLC unit cells on a printed circuit board, all-to-all coupled by switched capacitances. The experiments show Gaussian sampling and matrix inversion.
Where does the 10,000× energy figure come from?
arXiv:2510.23972, from Extropic with an MIT coauthor, says a simulated denoising thermodynamic model could match GPU Fréchet Inception Distance on binarized Fashion-MNIST while using about 10,000 times less energy. That figure is a system-level analysis, not a measured Z1 wall-plug result.
What analog-AI result did IBM Research publish?
The 23 August 2023 Nature paper reports a 14 nm chip with 35 million phase-change memory devices across 34 tiles and up to 12.4 tera-operations per second per watt of chip-sustained performance, including a software-equivalent keyword-spotting network.
Can the dataset be downloaded?
Yes. JSON is at https://thermocompute.dev/data/dataset.json, CSV is at https://thermocompute.dev/data/dataset.csv, and the ARD v1.0 manifest is at https://thermocompute.dev/.well-known/ard.json.