Hardware Validation Report · March 2026

132/100
Composite
Score

AriaOS v2.2.2 on NVIDIA Jetson AGX Orin 64GB.
Real models. Real inference. Zero cloud. Zero connectivity required.

↓ Download Full Report → Request Phase Zero Briefing
132.6
Composite Score
100 = Reference
48
Tokens / Sec
Llama 3.2 3B · Offline
28
Tokens / Sec
Mistral 7B · Offline
51°C
Peak Thermal
No Throttling
245
API Req / Sec
Sustained · 0 Errors
4.5
GB/s NVMe Read
Sequential
Real-World Inference

Validated. Offline.
No synthetic numbers.

Every inference number on this page was produced by running actual language models through Ollama on the Jetson hardware. No simulations. No theoretical throughput. The eval rate printed at the bottom of each run is what you see here.

Model Size tok/s
Llama 3.2 3B
2.0 GB
48.05
tok/s
Mistral 7B
4.4 GB
27.55
tok/s
Qwen3 8B (direct)
5.2 GB
24.83
tok/s
Qwen3 8B (reasoning)
5.2 GB
24.38
tok/s

All models run via Ollama on Jetson AGX Orin 64GB.
Zero network connectivity. Zero cloud dependency.
Validated March 15, 2026.

132.6
/ 100 · Composite Benchmark Score
GPU Compute FP32
150
GPU Compute FP16
150
CPU Performance
150
AriaOS Throughput
122.8
AriaOS API
101.9
Memory Bandwidth
73.4
Thermal Validation

Stable under
full load.

Five benchmark runs. Peak thermal of 51°C. Delta under 1°C across all sensors. The Orin's thermal design sustains MAXN performance without throttling — a critical requirement for sustained field deployment.

Sensor
Before After Delta
CPU-therm
50.8°C 51.4°C +0.6°C
GPU-therm
46.9°C 47.7°C +0.8°C
SOC0-therm
48.4°C 49.1°C +0.7°C
SOC2-therm
47.6°C 48.1°C +0.5°C
tj-therm (peak)
51.1°C 51.4°C +0.3°C
Tboard
39.0°C 40.0°C +1.0°C
Methodology

How we ran it.

Five independent benchmark runs across two days. All results timestamped and saved as JSON. Raw data available on request.

01
Hardware
NVIDIA Jetson AGX Orin 64GB Developer Kit. 2TB NVMe. Power mode: MAXN (maximum performance). jetson_clocks enabled. No active cooling beyond stock thermal solution.
02
Inference
Models run via Ollama 0.17.7 on JetPack 5.1.2 / CUDA 11.4. Identical prompt across all models. eval rate extracted from --verbose output. No network connectivity during testing.
03
GPU Benchmarks
PyTorch 2.1.0 matrix multiply operations at FP32 and FP16 precision across five matrix sizes. Results normalized to Orin 64GB manufacturer reference (100 = baseline).
04
API Performance
AriaOS backend running in standard mode (2 workers). Endpoints tested under simulated load: 5, 10, and 20 concurrent connections. 245 req/sec sustained over 5 seconds.
05
Thermal
Sensor readings captured before and after each benchmark run via tegrastats. All six primary thermal zones recorded. No throttling events detected across all runs.
06
Reproducibility
All benchmark results saved as timestamped JSON to NVMe. Five runs showed less than 2% variance on GPU compute. Raw JSON available on request for independent verification.
📄
AriaOS Hardware Validation Report — March 2026
ResilientMind AI LLC · SDVOSB | VOSB · CAGE: 14JQ9 · v2.2.2 · Jetson AGX Orin 64GB
↓ Download PDF Report
Full Report

Read the complete report.

12-page hardware validation report with full benchmark data, thermal profiles, stress test results, energy vertical validation, and governance evidence.

Get Involved

Three ways
to engage.

💬
Phase Zero Validation
90-day paid research engagement. We validate AriaOS against your specific hardware and operational environment.
🖥️
Hardware Testing
Submit your edge hardware for AriaOS compatibility testing. We benchmark, document, and return results with a formal report.
📋
Research Briefing
30-minute technical briefing on AriaOS architecture, DDIL capabilities, and defense/energy vertical applications.
Request a Briefing
All inquiries responded to within 48 hours.
Defense and federal partners prioritized.

ResilientMind AI LLC · SDVOSB | VOSB · CAGE: 14JQ9
Your information is never shared or sold.

✓
Request Received

We'll respond within 48 hours.
Defense and federal inquiries are prioritized.

ResilientMind AI LLC · SDVOSB | VOSB · CAGE: 14JQ9