Compute

On this page 8

Three tiers, with different jobs and very different economics.

TierJobSizing
On boardReal-time detection, precision landing, dock logicOnly when a decision must be made in flight
Ground stationFlight control, live find map, first-pass triageA laptop, or a tablet plus a radio
ProcessingStitching, model inference, prescription exportOne 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:

CapabilityWhy on board
Wildlife searchThe find map has to reach the ground crew now, not after landing
Bird deterrenceThe flock is leaving
Weed control, spot spraying variantReal-time nozzle control
Autonomous docksPrecision landing and the go/no-go logic
OptionComputePowerEU sourceUS sourceIndicative
Raspberry Pi 5CPU only5 to 12 WBerryBaseThe Pi Hut€80 to €120
Pi 5 plus AI Kit (Hailo-8L)13 TOPS8 to 15 WBerryBaseThe Pi Hut€150 to €200
Jetson Orin Nano Super~67 TOPS7 to 25 WAntratek, ReicheltSeeed, Arrow€250 to €350
Seeed reComputer J4012 (Orin NX 16 GB)~100 TOPS10 to 25 WAntratekSeeed€800 to €1,000
Luxonis OAK-DCamera plus accelerator in one5 WLuxonisLuxonis€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

StepEffect
Export to ONNX, then TensorRT or Hailo's compiler2 to 5x over naive PyTorch
INT8 quantisation with a calibration set2 to 4x again, usually under 1% accuracy loss
Right-size the input640 px inference on a 20 MP frame is a tiling decision, not a resize
Frame skip5 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

ItemWhyIndicative
Rugged laptop or tabletMission control, live map, sunlight readable€400 to €2,000
Telemetry radioRFD868x in the EU€250 to €350
LTE routerCorrections in, findings out€150 to €250
PowerEnough 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

JobBound byRough scale
PhotogrammetryCPU cores and RAM, GPU helps25 ha at 2.5 cm/px, roughly 600 frames, 30 to 60 min
Model inferenceGPUThousands of frames per hour
Model trainingGPU and VRAMHours to days, occasionally
Raster derivativesCPU and diskMinutes

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

OptionCostNotes
Own box: Ryzen 9 or Threadripper, 128 GB RAM, RTX 4090 or 5090, 4 TB NVMe€3,000 to €6,000 oncePays back against any per-hectare licence within one season
Used workstation plus one GPU€1,200 to €2,500Perfectly adequate. Buy RAM, not clock speed
Dedicated GPU server, hosted€150 to €600 per monthHetzner has GPU dedicated lines; check current models
Cloud GPU on demandPer hourGood 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

LayerHoldsSizing
HotThis week's flights, raw frames2 to 8 TB NVMe
WarmOrthomosaics and index rasters as Cloud Optimized GeoTIFF10 to 50 TB spinning disk or object storage
ColdRaw frames older than a seasonCheap 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.