Technology
NVIDIA Jetson Development
The NVIDIA Jetson family is our default compute for edge programs where inference throughput, deterministic latency and Linux ecosystem support all matter at once - autonomous mobile platforms, vision-based inspection cells, and connected devices that must think locally.
What it is
What Jetson is and when to reach for it.
The Jetson family runs from Nano (entry-level, 472 GFLOPS) through Xavier NX and AGX Xavier (mid-tier) up to Orin AGX (up to 275 TOPS INT8 for INT8 sparse workloads). All ship with the same JetPack stack - Ubuntu-based Linux, CUDA, TensorRT and DeepStream - so a team fluent on one Jetson can ship on another with modest ramp cost.
We reach for Jetson when the compute constraint is real inference throughput at deterministic latency: multi-camera vision cells, autonomous mobile platforms running perception and planning stacks, and connected devices with ML models substantial enough that MCU-class hardware would fail the deadline. We stay on Coral or Hailo when the constraint is ultra-low power at modest inference load, and we drop to MCU-class silicon when the model can be quantised aggressively enough.
Where we use it
NVIDIA Jetson Development in shipped programs.
- Autonomous zero-turn mower - Jetson-class compute running ROS 2 perception, planning and safety monitors on a battery-powered mobile platform
- Weld-inspection cells - Jetson running vision inference at line rate inside a 45 ms cycle budget
- Multi-camera edge classifiers for perimeter monitoring - Jetson with DeepStream pipelining several streams into one inference path
Patterns and gotchas
Patterns that work, and the ones that catch first-time users.
Design a carrier, do not use the dev-kit in production. Jetson modules (SoM form-factor) are what production designs use; the dev-kit is a great starting point but not a shippable product. We design custom carrier boards with the connectivity, power and thermals the product actually needs - and validate them against real workload profiles, not idle.
Thermal is often the real bottleneck. Jetson Orin AGX pulls up to 60 W at peak; even Nano at 10 W needs meaningful heatsinking. The thermal envelope decides your enclosure design and often your fan/no-fan choice - we plan it in week one, not after the first field prototype throttles.
Use TensorRT, do not deploy raw PyTorch. ONNX export followed by TensorRT engine building typically triples throughput on the same model. The engine has to be rebuilt per JetPack version and per module SKU, so we build a repeatable engine-build step into the CI/CD, not a manual step.
Related
How this fits our practices.
NVIDIA Jetson Development is a lens on our edge AI development practice. Adjacent tools and services: robotics, computer vision, embedded Linux, ROS 2 development.
Tell us what you’re building.
Send the constraint that worries you most - a latency budget, a power budget, a certification date. We’ll tell you straight whether we’re the right team.