About
Every operation has a physics that performs it best.
A matrix multiply is cheapest as light through an interferometer mesh. A random sample is cheapest as thermal noise in a magnetic junction. A reversible transform is cheapest on logic that recovers its charge instead of burning it. A spiking network is cheapest on silicon that only spends energy when a neuron fires.
General-purpose processors run all of these on one substrate and pay one price: roughly eight orders of magnitude above the thermodynamic floor. tei-fabric is the orchestration layer that ends that flat tax — it reads a workload, decomposes it into primitives from the Periodic Stack of Computation, and dispatches each primitive to the substrate whose physics performs it at the lowest joule cost.
The five substrate columns
Photonic
tei-d-photonic
Coherent light through a Mach-Zehnder interferometer mesh. Multiplication happens in amplitude; integration happens on a photodetector.
Owns: Dense linear algebra — matmul, convolution, attention products, unitary transforms.
~30 fJ per multiply-accumulate at the system level.
Stochastic
tei-d-stochastic
Probabilistic bits built on stochastic magnetic tunnel junctions. The thermal bath does the sampling — the compute is the noise.
Owns: Sampling — MCMC, Gibbs, simulated annealing, Bayesian posteriors.
~1 fJ per p-bit update.
Reversible
tei-d-reversible
Adiabatic CMOS recovers charge instead of dissipating it. The bijective phase of a computation approaches the Landauer floor.
Owns: Every primitive with a Bennett decomposition in the catalog — the reversible kernel runs nearly free; only the irreversible projection pays.
~10³ above the Landauer floor for the reversible phase.
Neuromorphic
tei-d-neuromorphic
Leaky integrate-and-fire neurons spend energy per spike event, not per clock. A silent neuron is nearly free.
Owns: Spiking networks — LIF inference, STDP learning, event-driven workloads.
~20 pJ per synaptic operation.
In-memory
tei-d-in-memory
Resistive crossbars compute matrix-vector products with Ohm's law and Kirchhoff's current law — the weights are conductances, the multiply is physics.
Owns: Matrix-vector multiply and small dense matmuls, at the array.
~2-5 fJ per multiply-accumulate on a 256×256 array.
Baseline
tei-d-baseline
A modern CPU/GPU, anchored to measured silicon. The comparator every other substrate is priced against.
Owns: everything, universally — at the general-purpose price.
~36 pJ per multiply-accumulate.
How the dispatcher works
Drop in an ONNX model — or build a workload by hand — and the importer maps each operator to a Periodic Stack primitive, resolving tensor shapes through the graph. The dispatcher then prices every primitive on every substrate from first principles: laser efficiency and ADC samples for the photonic column, crossbar tiling for the in-memory column, spike counts for the neuromorphic column, Bennett fractions for the reversible column. The lowest-joule supported substrate wins, and the full considered set is reported so you can see exactly why.
Every constant in every model carries a citation to the published literature — Landauer 1961, Shen 2017, Davies 2018, Wan 2022, the Murmann ADC survey, and the rest. The constants are also engineering knobs: drag the wavelength-channel count or the crossbar size and watch the cost surface re-route in real time.
Open by design
tei-fabric is Apache-2.0 Rust, built to compose the open substrate-simulation ecosystem — gdsfactory and SAX for photonics, thrml for thermodynamic sampling, CrossSim for crossbars, Lava and NIR for spiking networks, the DeBenedictis adiabatic-analysis flow for reversible CMOS. The fabric's job is the orchestration above all of them: joule-aware dispatch across columns those tools individually validate.
Built by Thermodynamic Edge Intelligence Corp. — a subsidiary of Open Interface Engineering. Sibling instruments: orbit.thermoedge.ai · tesilicon.thermoedge.ai · corridor.thermoedge.ai