NVIDIA Jetson Design Services
NVIDIA Jetson is a family of embedded AI compute modules — from Jetson Orin Nano to AGX Orin — that bring GPU-accelerated machine learning and computer vision to the edge.
Jetson is the choice when a product must run real neural networks locally — multi-camera vision, robotics autonomy, or generative/perception models — at the edge rather than in the cloud. The modules pair Arm CPUs with NVIDIA GPUs and run the same CUDA/TensorRT stack as the data center.
What Jetson actually buys you is software continuity. Models trained on data-center GPUs deploy through the same CUDA, cuDNN and TensorRT stack at the edge; JetPack supplies a maintained Ubuntu-based Linux with drivers, DeepStream for video-analytics pipelines and Isaac for robotics. The hardware model is a compute module — CPU, GPU, memory and storage on one board — mated to a carrier board that provides your product's specific connectors, power and I/O. Development typically starts on an NVIDIA developer kit and moves to a production module on a custom carrier once requirements freeze.
The honest caveats: Jetson modules operate in roughly 7 W to 60 W envelopes, so enclosure thermal design is a first-class engineering task, not an afterthought. Module and carrier costs put Jetson an order of magnitude above MCU-class silicon, which is only justified when the workload is genuinely GPU-shaped. And advertised TOPS are reached through disciplined model optimization — quantization, TensorRT engine building, pipeline tuning — not by copying a Python training script onto the module and hoping.
Key product families
Jetson Orin Nano
Entry edge AI (up to ~40 TOPS) for cameras and lighter inference workloads. It comfortably runs one to a few camera streams through modern detection and classification models, making it the default starting point for smart cameras, inspection stations and on-vehicle analytics where power and cost are constrained.
Jetson Orin NX
Mid-range performance in a compact module for robotics and multi-stream vision. Pin-compatible with the Orin Nano footprint at up to ~100 TOPS, it is the pragmatic upgrade path when a product outgrows the Nano — more streams, bigger models, or headroom for simultaneous perception and planning — without redesigning the carrier board.
Jetson AGX Orin
Up to ~275 TOPS for autonomous machines and heavy multi-sensor perception. The AGX class carries the memory bandwidth and I/O (multiple camera inputs, 10 GbE, PCIe) for full robot autonomy stacks, many-camera installations, and on-device experimentation with transformer and VLM-class models that smaller modules cannot hold.
Earlier modules: Nano, TX2, Xavier
A large installed base of products still ships on pre-Orin modules, and several are approaching end of life. We handle the migration engineering — carrier redesign or adaptation, JetPack version jumps, and TensorRT re-optimization — so fielded products move onto the Orin generation without a ground-up redesign.
How NVIDIA compares
Jetson vs Nordic / STM32 microcontrollers
These solve different layers of a product. MCUs deliver deterministic, millisecond-level control on milliwatts; Jetson delivers Linux-class AI perception on watts. Most Jetson-based machines we design still include an MCU — typically an STM32 — for motor control, power sequencing and safety supervision, because a Linux SoC should never be the only thing standing between a robot and a fault.
Jetson vs Espressif (ESP32)
An ESP32-S3 runs TinyML: wake words, threshold anomaly detection, small classifiers. Jetson starts where that ends — real-time video analytics, multi-camera fusion, large-model inference. A useful heuristic: if the workload is one low-rate sensor stream and a model under a few megabytes, try TinyML first; if it involves live video or anything transformer-shaped, budget for a Jetson.
Jetson vs custom NPU-based compute
At high volume with a fixed, well-characterized workload, a custom board around a dedicated NPU can undercut module economics — we have shipped exactly that for a warehouse-robotics platform where the BOM target ruled out a module. Jetson wins when workloads evolve, models grow, or time-to-market dominates. Having built both, our recommendation between them is grounded in shipped hardware, not vendor preference.
When to choose it
- On-device deep learning, computer vision, or sensor fusion that exceeds an MCU/NPU
- Robotics and autonomous machines needing high TOPS with deterministic latency
- Multiple camera or LiDAR streams processed locally
- Products whose models will keep evolving — the CUDA/TensorRT stack absorbs model changes without hardware respins
- Data-sensitive deployments where video must be processed on-premise rather than streamed to the cloud
Typical applications
Design considerations we handle
Custom carrier board design
Power sequencing to NVIDIA's design guide, high-speed signal integrity for PCIe, USB 3 and 10 GbE, MIPI CSI routing for camera inputs, M.2 expansion and rugged connectors — carrier boards are dense, controlled-impedance designs. Our PCB team designs up to 24-layer boards and runs pre-layout and post-layout SI analysis on every high-speed interface.
Thermal and power engineering
Every Jetson product is a thermal design project. We model the module's power profiles against the enclosure early, select between fansinks, heat pipes and fanless conduction designs, and validate with instrumented soak testing — so the module holds its performance mode in a hot enclosure instead of throttling in the field.
Camera and sensor bring-up
Multi-camera products need device-tree work, V4L2 driver integration, GMSL or MIPI link design, lens and ISP tuning, and time synchronization across sensors for fusion. We take responsibility for the whole imaging path, from connector pinout to frames arriving correctly timestamped in the application.
JetPack, TensorRT and deployment
We optimize client models with TensorRT and quantization to hit real-time targets, build DeepStream pipelines for video analytics, and set up production essentials: secure boot and fuse strategy, disk encryption, A/B image updates over the air, and fleet observability for devices in the field.
EMC and regulatory for high-speed boards
Gigahertz-class interfaces radiate, and Jetson products still need FCC and CE marks. We design for emissions compliance from the stack-up onward — shielding strategy, filtered I/O, careful clock management — and support the lab campaign through our compliance-testing practice.
What we've shipped on NVIDIA
5G fleet telematics with on-vehicle AI (Jetson Orin Nano)
A connected-fleet platform pairing 5G connectivity with a Jetson Orin Nano running 40 TOPS of on-vehicle perception — driver assistance analytics, V2X integration and CAN-bus vehicle data fused at the edge, engineered for automotive power transients and thermal extremes.
Digital-twin building gateways (Jetson AGX Orin)
Floor-level edge gateways built on the Jetson AGX Orin (275 TOPS) aggregating data from hundreds of custom ESP32-S3 sensor nodes and running occupancy and energy analytics locally, feeding a cloud digital-twin platform for a commercial smart-building deployment.
Warehouse robotics perception (custom edge-AI compute)
For a warehouse-robotics platform with hard BOM targets we shipped perception on a custom Cortex-A78 board with a dedicated 4 TOPS NPU instead of a module — experience on the alternative path that keeps our Jetson-versus-custom guidance honest when clients face the same trade.
How Rapid Circuitry designs with NVIDIA
Rapid Circuitry designs custom Jetson carrier boards — power sequencing, high-speed camera (MIPI CSI) and PCIe routing, thermal solutions, and rugged I/O — and integrates the sensors and firmware around the module. Our edge-AI work spans vehicles, buildings and robots for clients across the US, UK, EU, Australia and India, backed by a PCB practice that routinely delivers the dense, controlled-impedance boards (up to 24 layers) that Jetson carriers demand.
Frequently asked questions
When do I need a Jetson instead of a microcontroller?
Use a Jetson when the workload is real deep learning — multi-camera vision, large models, or robotics autonomy — that an MCU or small NPU cannot handle. For keyword spotting, simple anomaly detection or low-rate sensor ML, a microcontroller with TinyML is cheaper and lower power.
What does a Jetson carrier board do?
The Jetson module holds the processor and memory; the carrier board provides the connectors, power, camera and PCIe interfaces, and I/O your product needs. Most products need a custom carrier board designed to their form factor and sensors.
Orin Nano, Orin NX or AGX Orin — how do we choose?
Size the module to the perception workload with margin. One to a few camera streams through standard detection models: Orin Nano. Several streams, larger models, or simultaneous perception and planning: Orin NX — and its pin compatibility with the Nano lets one carrier design serve both. Many-camera autonomy, heavy sensor fusion or transformer-class models: AGX Orin. We benchmark the actual model on target hardware before locking the choice, because published TOPS figures and real pipeline throughput can differ substantially.
What does a custom Jetson carrier board cost, and how long does it take?
A carrier is a complex, high-speed, controlled-impedance design — as a typical range, expect $10,000–$25,000 for the board design effort, consistent with the complex-board tier in our PCB design cost guide, plus prototypes, bring-up and any camera or thermal engineering the product needs. Schedule-wise, 8–14 weeks from kickoff to a validated first article is typical. Many teams sensibly launch pilots on an off-the-shelf carrier and commit to custom hardware once requirements freeze.
Building on NVIDIA?
Tell us your requirements and we'll confirm the right part and design the board and firmware around it.
Start your design