VaidhyaMegha/edge-ai-kg

Edge AI Deployment Knowledge Graph

25,152 nodes. 76,306 edges. Boards, kernels and neural networks in one graph — so you can ask what actually runs on your silicon.

Built with Samyama Graph. Loader and generator: samyama-ai/edge-ai-kg.

Part real, part synthetic — and every node says which

Every node carries a provenance property ("real" or "synthetic") and a source. No node is unstamped:

provenance	Nodes
synthetic	23,910
real	1,242

Do not conflate them. The synthetic fleet exists to give the graph scale and realistic topology to query against; it is not a survey of deployed hardware. The real layer is measured fact from three public sources. Filter on provenance = 'real' before quoting any number as an observation about the world.

Sources — all three verified permissive

Source	Licence	Contributes
onnx/onnx — docs/Operators.md	Apache-2.0	Operator — name, domain, opset version
microsoft/onnxruntime — docs/OperatorKernels.md	MIT	Kernel, Accelerator — which execution provider implements which operator, at which opset range
mlcommons/tiny_results_v1.2 — summary.csv	Apache-2.0	Deployment, Board, SoC, Vendor, Runtime, BenchmarkTask, Model — measured throughput, accuracy and energy per inference on real commercial hardware

Each licence was checked at source rather than taken from documentation. All three permit redistribution; the MIT component is more permissive than Apache-2.0, so the dataset carries Apache-2.0 overall. The synthetic layer is generated by this project.

Nothing is held back — every node and edge loaded is published.

Why a graph

"Will this model run on this board?" is a join across vendor datasheets, runtime kernel tables and benchmark results that nobody maintains in one place. As a graph it is a path:

// Which operators has no execution provider implemented — the porting gap
MATCH (o:Operator)
WHERE NOT (:Kernel)-[:IMPLEMENTS]->(o)
RETURN o.name, o.domain ORDER BY o.name

// Real measured deployments only, ranked by energy per inference
MATCH (d:Deployment)-[:ON_BOARD]->(b:Board)
WHERE d.provenance = 'real'
RETURN b.name, d.energy_per_inference, d.throughput
ORDER BY d.energy_per_inference LIMIT 10

// Kernel coverage by execution provider
MATCH (k:Kernel)-[:PROVIDED_BY]->(r:Runtime)
RETURN r.name, count(k) AS kernels ORDER BY kernels DESC

Files

Nodes carry an id; edges reference those ids as src and tgt.

nodes/

File	Rows
nodes/kernel.csv	22,583
nodes/deployment.csv	1,513
nodes/operator.csv	377
nodes/modelvariant.csv	240
nodes/board.csv	134
nodes/accelerator.csv	91
nodes/model.csv	64
nodes/soc.csv	52
nodes/clinicaltask.csv	18
nodes/signalstage.csv	16
nodes/vendor.csv	15
nodes/sensor.csv	14
nodes/runtime.csv	13
nodes/dataset.csv	12
nodes/certification.csv	6
nodes/benchmarktask.csv	4

edges/

File	Rows	Connects
edges/implements.csv	22,583	Kernel -> Operator
edges/provided_by.csv	22,583	Kernel -> Runtime
edges/runs_on.csv	22,583	Kernel -> Accelerator
edges/on_board.csv	1,513	Deployment -> Board
edges/via_runtime.csv	1,500	Deployment -> Runtime
edges/uses_accelerator.csv	1,451	Deployment -> Accelerator
edges/of_variant.csv	1,440	Deployment -> ModelVariant
edges/uses_operator.csv	1,069	Model -> Operator; SignalStage -> Operator
edges/targets.csv	429	Runtime -> Accelerator
edges/variant_of.csv	240	ModelVariant -> Model
edges/made_by.csv	186	Board -> Vendor; SoC -> Vendor
edges/has_soc.csv	134	Board -> SoC
edges/certified_for.csv	104	Board -> Certification
edges/has_accelerator.csv	85	SoC -> Accelerator
edges/trained_on.csv	82	Model -> Dataset
edges/measures.csv	73	Deployment -> Model
edges/solves.csv	64	Model -> BenchmarkTask; Model -> ClinicalTask
edges/precedes.csv	60	SignalStage -> Model
edges/requires_sensor.csv	51	ClinicalTask -> Sensor
edges/next_stage.csv	40	SignalStage -> SignalStage
edges/governed_by.csv	22	ClinicalTask -> Certification
edges/feeds.csv	14	Sensor -> SignalStage

IMPLEMENTS, PROVIDED_BY and RUNS_ON are each 22,583 — one per kernel registration, which is the dense core of the graph.

Usage

from datasets import load_dataset
kernels = load_dataset("VaidhyaMegha/edge-ai-kg", "kernel", revision="v1.0")

Verification

• Totals reconcile against the snapshot header: 25,152 nodes, 76,306 edges.
• The loader verified its own writes: "verified: 25,152 nodes in graph 'default'". Load took 28.9s.
• 0 dangling edges. 0 orphan nodes.
• Every node is provenance-stamped — 1,242 real, 23,910 synthetic, none missing.
• Round-trip verified after upload.

Limitations

• Built from live upstream docs, so counts drift. This build came out at 25,152 / 76,306 against the repository's recorded 25,145 / 76,291: docs/OperatorKernels.md in ONNX Runtime gained kernel registrations between their measurement and this one (2026-08-29). Nothing is wrong — cite the revision.
• The synthetic fleet is generated, seed 20260814, scale 1.0. Its topology is plausible, not observed. Board/model combinations in the synthetic layer do not mean anyone shipped them.
• Operator grouping is this project's, not the ONNX standard's — defined in etl/onnx_catalog.py. Do not read it as an upstream taxonomy.
• MLPerf Tiny v1.2 is a small, specific benchmark — 73 submissions across 7 organisations, four benchmark tasks. It is not a survey of edge AI hardware performance generally.
• Kernel coverage reflects the ONNX Runtime documentation at fetch time, which lags the code.

Citation

Edge AI Deployment Knowledge Graph, v1.0 (25,152 nodes, 76,306 edges).
Built with Samyama Graph. https://huggingface.co/datasets/VaidhyaMegha/edge-ai-kg
Loader: https://github.com/samyama-ai/edge-ai-kg
Real sources: onnx/onnx (Apache-2.0); microsoft/onnxruntime (MIT);
              mlcommons/tiny_results_v1.2 (Apache-2.0).
Synthetic layer generated by this project, seed 20260814.
Retrieved: 2026-08-29.