Llama 3.2 3B Instruct on Jetson AGX Orin 64GB: 1 row.
Method v1.0 · dataset 2026-09-07 · verified 2026-09-07 · last updated September 2026
llamallama-3.2-3b-instruct on Jetson AGX Orin 64GB: 1 row (1 measured, 0 derived). Expected: INT4 80.4 tokens/s. Fastest: Jetson AGX Orin 64GB 80.4 tokens/s (class C), Jetson Orin Nano Super 43.07 tokens/s (class C), Jetson Orin Nano 8GB 27.7 tokens/s (class C).
Change hardware, precision, runtime or add a comparison set on the live engine.
Expected range per precision
Min / median / max of the matching rows, all classes. When a precision has no row, the estimator's benchmark hierarchy (B → C → D) supplies a fallback with its tier chain reported.
| Precision | Metric | N | Min | Median | Max | Best class | Measured |
|---|---|---|---|---|---|---|---|
| INT4 | tok/s | 1 | 80.4 | 80.4 | 80.4 | C | 1 |
Rows
| Precision | Runtime | Input | Power mode | Software | Metric | Class | Date | Source | Derivation | Dupes |
|---|---|---|---|---|---|---|---|---|---|---|
| INT4 | mlc | — b1 | default | — | 80.4 tok/s | C | 2026-09-07 | NVIDIA Jetson AI Lab (archive) · jetson-ai-lab:tutorial_slm.html · verified 2026-09-07 | 0 |
Cross-hardware comparison
Best row per hardware for the same model and precision, relative to the fastest.
| Hardware | Best value | Class |
|---|---|---|
| Jetson AGX Orin 64GB | 80.4 tok/s | C |
| Jetson Orin Nano Super | 43.07 tok/s | C |
| Jetson Orin Nano 8GB | 27.7 tok/s | C |
Assumptions
- Class C = measured and published by the source named on the row; class D = EdgeAIStack-derived (derivation stated). Rows are inference-only unless scope says end_to_end; publisher measurement conditions (JetPack, batch, warm-up) are in the linked page.
Method and limitations
Rows are read from the normalised benchmark database (publisher benchmarks and research-derived rows, deduplicated per hardware × model × runtime × precision × input × power mode, best evidence kept). Expected range = min / median / max of the matching rows per precision; when a precision has no row the estimator's benchmark hierarchy (B → C → D) is consulted and its tier chain reported. Comparison = best row per hardware for the same model and precision, relative to the fastest. Nothing is scored; every number carries its class and source line.
Full method, dedupe rule, evidence classes and the confidence rule: Benchmark Explorer methodology.
Measure it yourself.
Nothing measured for your exact config, or want to confirm a class-D row? Run the measurement protocol on your board and submit the result — a human checks it against the protocol before it becomes a class-C row with your attribution.
Links
- Live engine (this exact query): https://edgeaistack.ai/engines/benchmark-explorer/?hw=jetson_agx_orin&model=llama&variant=llama-3.2-3b-instruct&mv=1.0&dv=2026-09-07
- Methodology: /methodology/benchmark-explorer/
- Measurement protocol: /methodology/measurement-protocol/
- Dataset: /datasets/benchmarks-v2.json
- Platform page: Jetson AGX Orin 64GB
- Jetson AGX Orin 64GB overview: /benchmarks/jetson-agx-orin/
Change one input and re-run.
The live engine keeps every source line and gives you a fresh permalink and summary.