YOLO26n · FP32
Download this report as PDF — it is exactly what a client receives.
The verdict
It runs on all 3 configurations measured, with correct results. The fastest: Intel N100 · ONNX Runtime, at 57.6 ms per inference, about 17 per second.
Performance
Model inference, in milliseconds. “End to end” adds preparing the input and processing the output.
| CONFIGURATION | P50 | P95 | P99 | END TO END P50 | PER SECOND |
|---|---|---|---|---|---|
| Raspberry Pi 5 · 8 GB · ONNX Runtime 1.30.0 | 126.4 | 134.4 | 138.3 | 138.1 | 8 |
| Intel N100 · 8 GB · ONNX Runtime 1.30.0 | 57.6 | 58.9 | 59.3 | 67.2 | 17 |
| Intel N100 · 8 GB · OpenVINO 2026.4.0 | 64.0 | 65.6 | 71.3 | 72.4 | 16 |
Correct results
Outputs are compared with a reference computed with ONNX Runtime on a reference machine, with the same inputs.
- Raspberry Pi 5 · ONNX Runtime ✓ finds 100% of the reference objects, with 100% precision, across 12 images
- Intel N100 · ONNX Runtime ✓ finds 100% of the reference objects, with 100% precision, across 12 images
- Intel N100 · OpenVINO ✓ finds 100% of the reference objects, with 100% precision, across 12 images
What was expected and what happened
Before measuring, the protocol writes down what it expects in each dimension. Then it is checked against what was measured.
| DIMENSION | RASPBERRY PI 5 · ONNX RUNTIME | INTEL N100 · ONNX RUNTIME | INTEL N100 · OPENVINO |
|---|---|---|---|
| Valid ONNX model | ✓ passesonnx.checker OK | ✓ passesonnx.checker OK | — |
| Runtime load | ✓ passesload 57.68 ms | ✓ passesload 116.609 ms | ✓ passesload 359.373 ms |
| Operator support | ✓ passessession created with no rejected operators | ✓ passessession created with no rejected operators | — |
| Fits in memory | ✓ passespeak RSS = 3.0% of RAM | ✓ passespeak RSS = 3.3% of RAM | ✓ passespeak RSS = 3.5% of RAM |
| Correct results | ✓ passesrecall 100% · precision 100% vs reference | ✓ passesrecall 100% · precision 100% vs reference | ✓ passesrecall 100% · precision 100% vs reference |
| p95 ≤ 200 ms | ◆ meets itp95 134.3902 ms vs ≤ 200 ms | ◆ meets itp95 58.8727 ms vs ≤ 200 ms | ◆ meets itp95 65.5797 ms vs ≤ 200 ms |
| Thermal stability | ◆ passesTmax 65.3 °C | — | — |
| CPU support | — | — | ✓ passesfully executed on CPU |
| Faster than ONNX Runtime | — | — | ◆ nop50 64.045 ms vs 57.5655 ms in C08 (-11.3%) |
✓ the expectation was confirmed · ◆ it was unknown and is now measured · · not evaluated in this measurement
14 predictions confirmed, none wrong.
Not evaluated in this measurement: export.
Memory, load and temperature
| RASPBERRY PI 5 · ONNX RUNTIME | INTEL N100 · ONNX RUNTIME | INTEL N100 · OPENVINO | |
|---|---|---|---|
| Model load | 58 ms | 117 ms | 359 ms |
| Peak memory | 239 MB · 3.0% of RAM | 255 MB · 3.3% of RAM | 265 MB · 3.5% of RAM |
| Temperature | 57 → 65 °C | 53 → 62 °C | 54 → 64 °C |
| CPU frequency | 2400–2400 MHz | 2816–2900 MHz | 2821–3150 MHz |
| Thermal throttling | no | no, by frequency (no direct reading) | no, by frequency (no direct reading) |
Sustained test: 30 minutes non-stop
The hundred measurements above are on a cold device. Here the model runs for 30 minutes straight in one process, with ONNX Runtime's default threads and a random input of its size, to see how the device performs hot. Each row is 5 minutes; latency in milliseconds.
Raspberry Pi 5 · 8 GB
| MINUTES | P50 | P95 | MAX TEMP | REAL CLOCK MIN | FIRMWARE FLAGS |
|---|---|---|---|---|---|
| 0–5 | 128.7 | 134.5 | 79.0 °C | 2400 MHz | — |
| 5–10 | 128.8 | 134.7 | 82.3 °C | 2256 MHz | soft limit 5 s |
| 10–15 | 130.1 | 136.1 | 84.0 °C | 2201 MHz | soft limit 217 s |
| 15–20 | 131.1 | 137.0 | 84.0 °C | 2201 MHz | soft limit 273 s |
| 20–25 | 132.3 | 138.4 | 84.5 °C | 1500 MHz | soft limit 279 s, throttled 1 s, frequency capped 1 s |
| 25–30 | 133.3 | 139.7 | 85.1 °C | 1500 MHz | soft limit 284 s, throttled 2 s, frequency capped 2 s |
p50 went from 128.7 to 133.3 ms (+3.6%) between the first and the last 5 minutes. The firmware set the soft temperature limit at minute 7.8 and throttling at minute 17.5. Linux reported 2400 MHz; the processor's real clock, read from the firmware, went as low as 1500 MHz.
Intel N100 · 16 GB
| MINUTES | P50 | P95 | MAX TEMP | FREQUENCY MIN |
|---|---|---|---|---|
| 0–5 | 55.6 | 56.7 | 74.0 °C | 2763 MHz |
| 5–10 | 55.7 | 56.7 | 76.0 °C | 2762 MHz |
| 10–15 | 56.0 | 57.0 | 78.0 °C | 2761 MHz |
| 15–20 | 56.1 | 57.2 | 78.0 °C | 2759 MHz |
| 20–25 | 56.1 | 57.2 | 79.0 °C | 2761 MHz |
| 25–30 | 56.2 | 57.2 | 78.0 °C | 2770 MHz |
This run is from 2026-11-09, with 16 GB of memory; the hundred measurements above were taken with 8 GB. p50 went from 55.6 to 56.2 ms (+0.9%) between the first and the last 5 minutes. This device exposes no throttling flags: it shows in the temperature and the frequency the system reports.
The model
- File
- yolo26n.onnx · 9.5 MB · FP32
- Source
- Ultralytics yolo26n.pt (official; never a third-party conversion)
- License
- AGPL-3.0
- Inputs
- images FLOAT [1, 3, 640, 640]
- Outputs
- output0 FLOAT [1, 300, 6]
- Graph
- ONNX opset 17 · 384 nodes · 23 operator types
- File identity
- 843640bc7ae797dfbf3d9828ced4bfce393a78949b6ed282a0db874f64ef2680
- Graph identity
- f21c7ca23a529adaf73a497aba4b366aa0d94f9f5bdf02c55aebbbe1460bd90f
Measurement conditions
| RASPBERRY PI 5 · ONNX RUNTIME | INTEL N100 · ONNX RUNTIME | INTEL N100 · OPENVINO | |
|---|---|---|---|
| System | Debian GNU/Linux 13 (trixie) | Ubuntu 24.04.5 LTS | Ubuntu 24.04.5 LTS |
| Python | 3.13.5 | 3.12.3 | 3.12.3 |
| Effective precision | FP32, the model's own | FP32, the model's own | f32, pinned and verified |
| Threads | runtime defaults | runtime defaults | runtime defaults |
| Measurements | 100 after 20 warm-up runs | 100 after 20 warm-up runs | 100 after 20 warm-up runs |
| Before measuring | device freshly rebooted | device freshly rebooted | device freshly rebooted |
| Frozen environment | ENV-T1-RPI5-8GB-LAB01-v01 | ENV-T2-N100-8GB-LAB01-v02 | ENV-T2-N100-8GB-LAB01-v02 |
| Observation | OBS-C07-20261003T215041Z-b5c74f | OBS-C08-20261003T224756Z-1c9208 | OBS-C09-20261003T224811Z-9534d9 |
What this report does not cover
- Power draw: it was not measured.
- Your enclosure, your airflow and your ambient temperature: the devices sit in an office, with their heatsinks.
- The sustained test with OpenVINO: it is done with ONNX Runtime only.
- Integrated GPUs and accelerators: only the CPU was measured.
- More than one input at a time: everything is batch 1, measuring latency.
- Other runtimes, and devices other than the ones shown here.
This is a sample report. The model is the official Ultralytics YOLO26n, which is public, and that is why this report can be published. Client reports are confidential, and the model is deleted when they are delivered.
Generated on 9 November 2026.