Compute
On this page 8
Three tiers, with different jobs and very different economics.
| Tier | Job | Sizing |
|---|---|---|
| On board | Real-time detection, precision landing, dock logic | Only when a decision must be made in flight |
| Ground station | Flight control, live find map, first-pass triage | A laptop, or a tablet plus a radio |
| Processing | Stitching, model inference, prescription export | One GPU box. This is the one that replaces a licence |
On board
Most capabilities do not need it. A prescription that gets loaded before the next pass is fine processed on the ground within the hour, and every on-board computer adds weight, heat, power draw and a failure mode.
Add it when the answer has to exist during the flight:
| Capability | Why on board |
|---|---|
| Wildlife search | The find map has to reach the ground crew now, not after landing |
| Bird deterrence | The flock is leaving |
| Weed control, spot spraying variant | Real-time nozzle control |
| Autonomous docks | Precision landing and the go/no-go logic |
| Option | Compute | Power | EU source | US source | Indicative |
|---|---|---|---|---|---|
| Raspberry Pi 5 | CPU only | 5 to 12 W | BerryBase | The Pi Hut | €80 to €120 |
| Pi 5 plus AI Kit (Hailo-8L) | 13 TOPS | 8 to 15 W | BerryBase | The Pi Hut | €150 to €200 |
| Jetson Orin Nano Super | ~67 TOPS | 7 to 25 W | Antratek, Reichelt | Seeed, Arrow | €250 to €350 |
| Seeed reComputer J4012 (Orin NX 16 GB) | ~100 TOPS | 10 to 25 W | Antratek | Seeed | €800 to €1,000 |
| Luxonis OAK-D | Camera plus accelerator in one | 5 W | Luxonis | Luxonis | €200 to €400 |
Budget the whole chain, not the board: a Jetson at 20 W for 30 minutes is 10 Wh off the flight battery, which is a couple of minutes of endurance, plus a heat sink and airflow that a sealed payload bay does not have.
Making a model fit
| Step | Effect |
|---|---|
| Export to ONNX, then TensorRT or Hailo's compiler | 2 to 5x over naive PyTorch |
| INT8 quantisation with a calibration set | 2 to 4x again, usually under 1% accuracy loss |
| Right-size the input | 640 px inference on a 20 MP frame is a tiling decision, not a resize |
| Frame skip | 5 fps is plenty for a search at 5 m/s |
A quantised small detector runs comfortably at 15 to 30 fps on a Pi with the AI Kit, which is more than a wildlife search needs.
Ground station
| Item | Why | Indicative |
|---|---|---|
| Rugged laptop or tablet | Mission control, live map, sunlight readable | €400 to €2,000 |
| Telemetry radio | RFD868x in the EU | €250 to €350 |
| LTE router | Corrections in, findings out | €150 to €250 |
| Power | Enough for a full morning | €100 |
Software: QGroundControl or Mission Planner will do everything at the start. Write your own planner when the mission patterns become capability-specific, which happens around the point you are flying the same three route templates every week.
Processing
This is the box that replaces the per-hectare licence, and it is the best value purchase in the whole platform.
What the work actually is
| Job | Bound by | Rough scale |
|---|---|---|
| Photogrammetry | CPU cores and RAM, GPU helps | 25 ha at 2.5 cm/px, roughly 600 frames, 30 to 60 min |
| Model inference | GPU | Thousands of frames per hour |
| Model training | GPU and VRAM | Hours to days, occasionally |
| Raster derivatives | CPU and disk | Minutes |
Photogrammetry is memory hungry. 64 GB is a sensible floor for real fields and 128 GB removes a whole class of failures.
Build or rent
| Option | Cost | Notes |
|---|---|---|
| Own box: Ryzen 9 or Threadripper, 128 GB RAM, RTX 4090 or 5090, 4 TB NVMe | €3,000 to €6,000 once | Pays back against any per-hectare licence within one season |
| Used workstation plus one GPU | €1,200 to €2,500 | Perfectly adequate. Buy RAM, not clock speed |
| Dedicated GPU server, hosted | €150 to €600 per month | Hetzner has GPU dedicated lines; check current models |
| Cloud GPU on demand | Per hour | Good for training bursts, expensive as a steady state |
The sane pattern: own the steady state, rent the bursts. One box handles daily flights; rent cloud GPU for the week you retrain a model.
Storage
| Layer | Holds | Sizing |
|---|---|---|
| Hot | This week's flights, raw frames | 2 to 8 TB NVMe |
| Warm | Orthomosaics and index rasters as Cloud Optimized GeoTIFF | 10 to 50 TB spinning disk or object storage |
| Cold | Raw frames older than a season | Cheap archive, or delete after the derived products are verified |
One 25 hectare flight at 2.5 cm/px is roughly 600 frames at 25 MB, so about 15 GB raw, and 1 to 3 GB of derived products. A busy operator flying 40 fields a week generates several terabytes a season. Decide the raw frame retention policy on day one, in writing, and put it in the customer contract.
Object storage: an S3-compatible store keeps the code portable. SeaweedFS (Apache-2.0) and Garage are the permissive self-hosted options; MinIO is AGPL-3.0, which matters if you modify it. See Platform.
Where this repository sits
The site and the API described in the guide run as a tenant on a shared box and want to answer in milliseconds. The processing pipeline wants a GPU and hours of wall clock. Keep them apart: the only contract between them is the flight record. See Deployment.