NVIDIA Jetson
NVIDIA Jetson is a family of compact computer modules and developer kits with NVIDIA GPUs for running AI inside robots, drones, cameras and other edge devices. It spans entry modules up to the Blackwell-based Jetson Thor series and runs the JetPack software stack.1
Also known as Jetson Thor, Jetson AGX Orin, Jetson Orin NX, Jetson Orin Nano, Jetson modules and developer kits
At a glance
- What is it?
- Jetson is NVIDIA's line of embedded AI computers. Each Jetson is a small system-on-module that combines an Arm CPU, an NVIDIA GPU and memory, sold as a module for product designs or as a developer kit for prototyping. The current range includes the Jetson Thor series (Blackwell GPU), the Jetson Orin series (AGX Orin, Orin NX and Orin Nano) and older Xavier, TX2 and Nano modules. All run JetPack, NVIDIA's software stack built on Jetson Linux, with CUDA, cuDNN and TensorRT.12
- What does it do?
- Jetson runs AI models where the data is created instead of sending it to a data center. Typical workloads are camera analytics, robot perception and control, and local generative AI such as vision language models. JetPack supplies the drivers, AI libraries and frameworks (PyTorch, vLLM, SGLang, Dynamo-Triton), plus security features such as secure boot, disk encryption and over-the-air updates. NVIDIA SDKs such as DeepStream, Isaac ROS and Holoscan run on top.2
- Who needs it?
- Teams building robots, drones, smart cameras, medical or industrial devices that must run AI on board with limited power and space, or without a reliable network link. Developers prototyping edge AI on a developer kit before moving to a production module.
- What does it need?3
- A Jetson developer kit or module with a suitable carrier board, power supply and storage
- For the Orin Nano developer kit: a host PC with 25 GB free, a 16 GB+ USB drive, and a 64 GB+ microSD card or NVMe SSD
- JetPack 6.x-generation UEFI firmware on the Orin Nano kit before installing JetPack 7.2
- Linux and embedded development skills; ROS 2 skills for robotics
- Models optimized for the target memory and power budget (for example with TensorRT)
- A power and thermal budget that matches the chosen module
- What it is not
- Jetson is not a GPU card for a PC or a data center accelerator; it is an embedded computer that needs a carrier board in a product. It is not a microcontroller, so it uses far more power and costs more than one. A developer kit is meant for development; production designs use Jetson modules, often on partner carrier boards. For industrial-grade functional safety and enterprise software support, NVIDIA points to its separate IGX platform.1
Availability and licensing. NVIDIA's developer site states Jetson Thor is available to order. New T3000 and T2000 Thor modules have been announced, and the Jetson Orin Nano 2 page shows a Notify Me option rather than ordering. The product page lists the Jetson Orin Nano series starting at 199 USD. JetPack 7 supports Orin and Thor.14
The problem it solves
Many AI applications cannot wait for a round trip to the cloud. A robot must react within milliseconds, a camera on a remote site may have little bandwidth, and video from a factory or clinic may not be allowed to leave the building.
Running modern models on the device needs GPU-class compute in a small, low-power package with a maintained software stack. Jetson provides that hardware, and JetPack keeps the drivers, AI libraries and security features aligned across the module range.
How it works
- Module. A Jetson module holds the CPU, GPU and memory; it sits on a carrier board that provides connectors for cameras, networking and storage.
- Jetson Linux. The board support package (Ubuntu based, with Yocto Project support in JetPack 7.2) is flashed to the device, now also through a Jetson ISO installer.
- JetPack AI stack. CUDA, cuDNN and TensorRT accelerate models; PyTorch, vLLM, SGLang and Dynamo-Triton (formerly Triton Inference Server) serve them.
- SDKs and models. DeepStream for video, Isaac ROS for robots and Holoscan for sensor streams run on top, along with open models such as Nemotron, Cosmos and Isaac GR00T.
- Fleet operation. Secure boot, disk encryption, fTPM and over-the-air updates keep deployed devices maintained.12
Diagram as a list
Applications & solutions
- DeepStream, Isaac ROS, HoloscanDomain SDKs for video, robots and sensors
Models & frameworks
- Open models (Nemotron, Cosmos, GR00T)AI models deployed to the device
Inference & runtime software
- JetPack AI stack (CUDA, cuDNN, TensorRT)Accelerates model inferenceConnects to PyTorch, vLLM, SGLang, Dynamo-Triton, DeepStream, Isaac ROS, Holoscan
- PyTorch, vLLM, SGLang, Dynamo-TritonRuns and serves modelsConnects to Open models (Nemotron, Cosmos, GR00T)
Operations & orchestration
- Jetson Linux BSPOS, drivers, flashing and Yocto buildsConnects to JetPack AI stack (CUDA, cuDNN, TensorRT)
- Security and OTASecure boot, encryption and remote updatesConnects to Jetson Linux BSP
Accelerated computing
- Jetson module (Thor, Orin)CPU, GPU and memory on one boardConnects to Jetson Linux BSP
- Carrier board or developer kitConnectors for cameras, network and storageConnects to Jetson module (Thor, Orin)
Capabilities
Module range by performance and power1
From Jetson Orin Nano (up to 67 TOPS, 7 to 15 W) and Orin NX (up to 157 TOPS, 10 to 25 W) to AGX Orin (up to 275 TOPS, 15 to 60 W) and the Thor series (40 to 130 W).
Why it matters: Lets a team pick compute that fits the device's power and size limits.
Limits: TOPS figures are vendor peak numbers; real throughput depends on the model and precision.
Jetson Thor for high-end physical AI1
The Jetson AGX Thor Developer Kit has a Blackwell GPU and 128 GB of memory, and NVIDIA states up to 2070 FP4 TFLOPS (sparse) within 130 W, and up to 7.5x the AI compute and 3.5x the energy efficiency of Jetson AGX Orin.
Why it matters: Runs larger generative models and humanoid robot stacks on the device.
Limits: Both figures are vendor statements; the FP4 sparse number is a peak value.
JetPack 7 software stack2
Built on Ubuntu 24.04 and Linux kernel 6.8 with a preemptable real-time kernel, Multi-Instance GPU and Holoscan Sensor Bridge, supporting Orin and Thor; Thor uses a unified CUDA 13.0 install.
Why it matters: One maintained OS and AI stack across current modules.
Limits: Older modules (Xavier, TX2, Nano) are not listed as JetPack 7 targets; check the release for your module.
AI libraries and serving frameworks2
CUDA, cuDNN and TensorRT, plus PyTorch, vLLM, SGLang and Dynamo-Triton (formerly Triton Inference Server).
Why it matters: Lets the same models and serving tools used in the data center run on the device.
Limits: Large models must still fit the module's memory.
Security and over-the-air updates2
Jetson Linux includes secure boot, disk encryption, runtime integrity, fTPM and secure OTA updates.
Why it matters: Needed to keep devices in the field patched and tamper-resistant.
Limits: We recommend planning secure boot, encryption and OTA setup as part of product design and following the Jetson Linux security documentation.
Supported NVIDIA SDKs2
DeepStream for video analytics, Isaac ROS for robotics and Holoscan for real-time sensor processing run on Jetson.
Why it matters: Shortens development for common edge workloads.
Limits: Each SDK has its own JetPack version requirements.
Custom Linux and agent skills2
JetPack 7.2 officially supports Yocto Project builds, and Jetson agent skills help with BSP bring-up for custom carrier boards, memory optimization and model benchmarking; NemoClaw can be installed with one command.
Why it matters: Speeds the move from developer kit to a custom product image.
Practical use cases
Camera analytics must run on site because bandwidth is limited or video may not leave the premises.
- Approach
- Run a DeepStream or Metropolis VSS pipeline on a Jetson Orin or Thor device next to the cameras.
- Role of NVIDIA Jetson
- Provides the on-site GPU compute and the JetPack stack.
- Data, infrastructure and skills
- Camera streams, a module sized for the number of streams, and DeepStream skills.
- Type of benefit
- Data stays local
- Caveats
- Stream counts per module depend on resolution and models; test on the target hardware.
- First step
- Run a DeepStream sample on a developer kit with one of your cameras.
A mobile robot needs on-board perception and navigation.
- Approach
- Deploy Isaac ROS packages on a Jetson module inside the robot.
- Role of NVIDIA Jetson
- Acts as the robot's main AI computer.
- Data, infrastructure and skills
- ROS 2 skills, sensors, and a power budget that fits the module.
- Type of benefit
- Lower reaction latency
- Caveats
- Thermal design and power must be planned early in the robot design.
- First step
- Follow the Isaac ROS quick start on a Jetson developer kit.
A device should answer questions about what its camera sees without a cloud connection.
- Approach
- Run a small vision language model, for example from Jetson AI Lab tutorials or Cosmos3-Edge, on Jetson Orin or Thor.
- Role of NVIDIA Jetson
- Runs generative AI locally.
- Data, infrastructure and skills
- A module with enough memory for the chosen model.
- Type of benefit
- Offline operation
- Caveats
- Smaller models answer less reliably than large ones; validate on your scenes.
- First step
- Try a Jetson AI Lab tutorial on a developer kit.
Who uses it
Instacart (Maplebear Inc.) · Grocery retail technology
Instacart: Caper smart carts on Jetson and GPU ranking with Dynamo
Instacart runs item recognition on its Caper smart carts with NVIDIA Jetson Orin NX modules and moved online ad and item ranking to NVIDIA GPUs with Dynamo. Published results include 65 percent lower whole-page ranking latency and an incremental sales lift above 1 percent in A/B tests.
Scaling
Foxconn (Hon Hai Technology Group) · Electronics manufacturing
Foxconn: digital twins for new server plants with Omniverse, Isaac and Metropolis
Foxconn uses NVIDIA Omniverse digital twins to plan production lines, Isaac to simulate robots and Metropolis for camera-based monitoring, from Hsinchu to new server plants in Mexico and the US. Most published results are expectations, such as a forecast energy cut of over 30 percent in Mexico.
In production
Works with
Optional integration
- NVIDIA DeepStream SDKDeepStream is a supported SDK on JetPack.
- NVIDIA Holoscan SDKHoloscan SDK is a supported SDK on JetPack.
Complementary tools
- NVIDIA IsaacIsaac ROS and GR00T run on Jetson for on-robot deployment.
- NVIDIA MetropolisJetson AGX Thor is a validated edge platform for the VSS blueprint.
- NVIDIA CUDA ToolkitNVIDIA documents CUDA upgrades for Jetson devices; the toolkit targets embedded systems.
Optional integration for
- NVIDIA Dynamo-TritonNVIDIA publishes Triton binary releases for Jetson JetPack on GitHub.
- NVIDIA Holoscan SDKSupported on Jetson AGX Thor, Jetson AGX Orin and Orin Nano.
- NVIDIA Nsight Developer ToolsNsight tools are packaged in JetPack and Nsight Systems supports Jetson.
- NVIDIA DeepStream SDKDeepStream 9.1 packages target Jetson Orin and Jetson Thor on JetPack 7.2.
- NVIDIA AI WorkbenchJetson devices can be added as remote locations through a CLI install.
Relationship labels follow NVIDIA's documentation. "Alternative approaches" does not mean one is better: each profile says when it fits.
Getting started
Choose a developer kit
Compare module specifications (TOPS, power, memory) with your workload; the Orin Nano Super kit is the entry option and AGX Thor the high end.
Check: Expected model size and stream count fit the chosen module.
Confirm firmware (Orin Nano)
JetPack 7.2 needs JetPack 6.x-generation UEFI/QSPI firmware; older kits follow the JetPack 6.x update path first.
Check: Firmware version is on the 6.x generation or later.
Create the installer USB
Download the JetPack 7.2.1 Jetson ISO and write it to a USB drive of 16 GB or more (not to a microSD card).
Check: The kit boots from the USB installer.
Install Jetson Linux
Install onto a 64 GB+ microSD card or an NVMe SSD and complete the initial setup.
Check: The kit boots into Jetson Linux from its own storage.
Run a first workload
Install the JetPack components and run a DeepStream, Isaac ROS or Jetson AI Lab sample.
Check: The sample runs with GPU acceleration.
Official resources
Could this technology help you?
Describe your project to the Solution Architect. It starts with NVIDIA Jetson as context but recommends independently, including when you do not need it.
Sources
Each statement above links to the source it comes from. Labels say who reported it.
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.