The Rust runtime, simulation, and visualization for the Thermodynamic Edge Intelligence compute fabric.
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. tei-fabric reads a workload, decomposes it into primitives from the Periodic Stack of Computation (258 primitives, 33 families), and dispatches each primitive to the substrate whose physics performs it at the lowest joule cost.
Live demo: fabric.thermoedge.ai
| Dialect | Physics | Anchor | Key citations |
|---|---|---|---|
tei-d-baseline |
Modern CPU/GPU (the comparator) | ~36 pJ/MAC | Landauer 1961 · Bérut 2012 |
tei-d-photonic |
MZI mesh, WDM, modulators + ADC | ~30 fJ/MAC | Shen 2017 · Wang 2018 · Hamerly 2019 |
tei-d-in-memory |
RRAM crossbar MVM, tiled | ~2-5 fJ/MAC | Wan 2022 · Khaddam-Aljameh 2022 · CrossSim |
tei-d-stochastic |
sMTJ p-bit thermodynamic sampling | ~1 fJ/sample | Camsari 2017 · Borders 2019 · thrml |
tei-d-reversible |
Adiabatic CMOS + Bennett decomposition | ~10³ × Landauer (L₀ phase) | DeBenedictis 2020 · Bennett 1989 |
tei-d-neuromorphic |
LIF spike events + STDP | ~20 pJ/SOP | Davies 2018 · Merolla 2014 · NIR |
Every constant carries a citation. Every constant is also an engineering knob — the API accepts per-request substrate_params, and the web UI exposes them as live sliders with physical diagrams.
- ONNX import (
tei-import) — pure-Rust protobuf decode + fixed-point shape inference with constant folding. Handles CNNs (MobileNetV2), quantized transformers (BERT), and diffusion UNets (Stable Diffusion v1, 1.6 GB) including ORT contrib fused ops (FusedMatMul,MultiHeadAttention,GroupNorm, …). - Cost-surface dispatch (
tei-cost-surface) — per-invocation lowest-joule substrate selection with the full considered set reported. - HTTP service (
tei-serve) — Axum: batch dispatch, SSE streaming dispatch, chunked uploads (to 2.5 GB), catalog-driven workload presets. - Browser viz (
web/) — Astro: Sankey flow of primitives → substrates, live engineering-parameter sweeps, per-invocation cards.
tei-fabric/
├── crates/
│ ├── tei-stack/ Periodic Stack catalog loader + index
│ ├── tei-ir/ workload IR (matmul / sampling / spiking profiles)
│ ├── tei-substrate-traits/ the Substrate trait + Cost
│ ├── tei-d-*/ six substrate dialects (see table)
│ ├── tei-cost-surface/ dispatcher + presets + summaries
│ └── tei-import/ ONNX → Workload (protobuf + shape inference)
├── bin/tei-serve/ Axum API
├── data/stack.json Periodic Stack catalog (mirror of compute.openie.dev)
└── web/ browser viz (Astro 5 + Tailwind 4)
cargo test --workspace # physics anchors + routing + ONNX fixtures
cargo run -p tei-serve # API on :9651
cd web && pnpm install && pnpm dev # viz on :4321Requires protoc (protobuf compiler) for the vendored ONNX schema: brew install protobuf / apt install protobuf-compiler.
GET /health liveness + catalog size
GET /api/stack full Periodic Stack catalog
GET /api/substrates registered dialects + citations
GET /api/presets catalog-driven workload presets
POST /api/dispatch Workload [+ substrate_params] → DispatchPlan
POST /api/dispatch/stream same, as SSE (started/invocation/complete)
POST /api/import/onnx single-shot ONNX import (small models)
POST /api/import/onnx/chunk chunked import (X-Upload-Id / X-Chunk-Index / X-Chunk-Total)
Apache-2.0. See LICENSE-APACHE.
The Periodic Stack catalog (data/stack.json) is content from compute.openie.dev and tracks that source. tei-fabric composes the open substrate-simulation ecosystem (gdsfactory, SAX, thrml, CrossSim, Lava, NIR, the DeBenedictis adiabatic-analysis flow) rather than reinventing it — the fabric is the joule-aware orchestration above tools those communities validate.
Built by Thermodynamic Edge Intelligence Corp. — a subsidiary of Open Interface Engineering.