Export targets¶
| Target | Edition | What you get | Runs on |
|---|---|---|---|
| ONNX Runtime (CPU) | Standard | One self-describing .onnx; optional .zip with classes.txt, metadata.json, README |
ctrlX CORE or IPC CPU |
| Hailo accelerator | Pro | Bundle for the Hailo Dataflow Compiler: INT8 calibration set from your images, model script, decode data (metadata.json, dgqp.npz). Tools ▸ Compile Hailo Bundle to .hef… produces the .hef |
ctrlX CORE X7 with Hailo-8 or Hailo-8L |
| NVIDIA GPU (CUDA / TensorRT) | Pro | Graph ready for trtexec (no NMS, batch 1), plus run_nvidia.py as a worked example |
ctrlX IPC with NVIDIA GPU |
- Hailo is proven end to end on a ctrlX CORE X7 with Hailo-8. Parcel models with package corners also run on Hailo.
- NVIDIA on the ctrlX IPC: on-device demonstration in progress.
Measured result
On a ctrlX CORE X7, a Model Bench parcel model runs in 13.05 ms per frame on the Hailo-8 accelerator, against 246 ms on the X7's CPU — 19 times faster.
The Export ONNX dialog¶
Export ONNX (for ctrlX)… on the Train Model page opens one dialog for all three targets. Pick the target first: the other options follow from it.

For Hailo, NMS and variable batch size are switched off and locked, a calibration set is taken from your images, and the dialog explains what still has to happen on the device side.

For NVIDIA GPU, the graph is written without NMS and with batch 1, ready for trtexec.

Screenshots to come
- S20 — Compile Hailo Bundle: preflight checks or compile in progress.
- S21 — Visual Detector web UI on the ctrlX device with live detections from a Model Bench model.