NVIDIA Jetson Modules Compared: Orin Nano vs Orin NX vs AGX Orin vs Thor

NVIDIA Jetson modules are the system-on-modules at the heart of every Jetson computer, and the current range runs from the 7 W Orin Nano to the Jetson T5000 at up to 130 W. The right module is the smallest one that covers your AI workload inside your power, I/O and lifecycle budget, not the one with the biggest number.

In short: choose the Jetson Orin Nano for entry-level, low-power vision on one or two cameras; the Jetson Orin NX for multi-camera AI and heavier inference; the Jetson AGX Orin for demanding multi-sensor workloads; and Jetson Thor for generative and physical AI on humanoids and autonomous machines.

This guide compares all nine current modules side by side with NVIDIA’s published figures, then walks through how to size a module, how Thor compares with AGX Orin, which modules share a carrier board, how industrial and commercial modules differ, and what we have shipped. For the platform itself, start with what NVIDIA Jetson is.

The table lists every current module with NVIDIA’s published figures. Orin Nano and Orin NX are shown in their Super modes from JetPack 6.2.

Orin figures are sparse INT8 TOPS and Thor figures are sparse FP4 TFLOPS, so the columns are not directly comparable; NVIDIA rates Thor at up to 7.5 times the AI compute of AGX Orin. Our explainer on TFLOPS and TOPS covers why precision changes the figure, and the Orin Nano Super paper explains the Super modes behind the Orin Nano and Orin NX ratings.

Across the Orin family the range is roughly eightfold, from 34 to 275 TOPS, on one architecture and one software stack. That is why a project can often prototype on one Orin module and ship on another. Our earlier comparison of NVIDIA Orin for edge AI performance looks at the Orin modules in more depth.

Interfaces often decide the module as much as the AI figure. These are module capabilities; the carrier board decides which reach a connector.

Two points stand out. The Orin Nano has no hardware encoder, so recorders and video gateways usually start at Orin NX; our Orin Nano vs Orin NX explainer covers that in detail. And Thor moves camera-heavy designs towards Ethernet: its multi-gigabit links and NVIDIA’s Holoscan sensor bridge stream sensor data straight into GPU memory. Interface counts differ between T4000 and T5000, so check NVIDIA’s published specifications for each Jetson module for exact figures.

Start from the workload, then check memory, power and interfaces. The table is a first pass; benchmarks on the real model settle the choice.

Four worked examples

A sealed outdoor sensor running a single detection model needs little compute and no fan: an Orin Nano at 10 to 15 W fits a passively cooled, sealed enclosure, with Super headroom where the cooling allows.

A four-camera autonomous mobile robot fusing vision and navigation in real time is an Orin NX job: enough headroom for several concurrent pipelines without the size, power and cost of the largest module.

An agricultural vehicle with eight cameras detecting crops and weeds at speed suits an AGX Orin 64GB, with 16 camera lanes, two video encoders and the memory to hold several models.

A humanoid robot running perception, planning and a vision language action model together is where Jetson Thor earns its place. Each step up the range adds heat to remove, so keeping a Jetson module inside its thermal limits is part of the sizing decision rather than an afterthought.

Every module has a set of NVIDIA power modes, and the mode you ship decides the cooling, the power supply and the temperature the system can run at. Catalogue ratings show how much the mode matters.

System power is always above module power, because the carrier, storage, cameras and USB devices draw on top. Size the supply for the top mode plus everything else, then confirm under the real workload at the hottest ambient that the module holds its clocks. The Premio JCO-1000-ORN figures show the trade clearly: the same fanless chassis gives up 20 °C of operating range when the Orin NX moves from 25 W to 40 W.

For batteries and vehicles, the lower fixed modes are often the better choice. They keep power predictable, extend run time and leave thermal margin, at the cost of some peak throughput; our guide to thermal throttling explains what happens when a module runs out of margin instead.

Which Jetson module is best for robotics?

It depends on the robot. Small AMRs, delivery robots and research platforms with two to four cameras usually run well on Orin NX, which balances compute, memory and power for a battery-powered base. Outdoor robots and agricultural machines with six or more cameras and lidar move to AGX Orin. Humanoids and mobile manipulators running vision language action models move to Thor. Our robot computer guide sizes the compute by robot class and covers the real-time and safety controllers around it.

Jetson Thor is the generation after Orin, and its two modules sit well above AGX Orin. Our white paper on how Jetson Thor advances edge AI beyond AGX Orin covers the architecture; this is the side-by-side.

Thor is not a drop-in upgrade. It uses a new connector, a higher power envelope and new I/O, so it needs one of the Jetson Thor carrier boards rather than an AGX Orin board, and applications built on JetPack 6 are rebuilt for JetPack 7. For Orin fleets, what JetPack 7.2 and supported Yocto builds mean for production Orin fleets explains how the two software lines now meet. Our answer on whether Jetson Thor needs a new carrier board sets this out, and Jetson T5000 vs T4000 compares the two Thor modules.

The T4000 and T5000 share a form factor, so a Thor carrier is designed for both. Memory is usually what separates them: 128 GB on the T5000 allows larger models, longer context or several models side by side. NVIDIA also brings Multi-Instance GPU partitioning to Thor, which lets one module divide its GPU between isolated workloads. For classic detection and tracking on a few cameras, AGX Orin remains a lower-power and well-proven choice; our comparison of Orin NX vs AGX Orin covers the step below it.

Jetson modules fall into a few connector families, and moving within a family is far easier than moving between them. The usual migration paths:

The previous generation is still in many fielded systems. The Jetson Xavier family and the older TX2 have reached an accelerated end of life, and our guide to Xavier and TX2 end-of-life and migration to Orin covers last-time-buy and refresh planning. For the AGX step specifically, AGX Orin vs AGX Xavier sets out the generational gain. The software stack carries across, so the application generally moves with it after a rebuild.

Shared footprints only help if the carrier’s power and software keep up. Our guide to choosing a Jetson carrier board covers what to look for, and the explainer on SO-DIMM covers the connector the smaller modules use.

Most Jetson modules are commercial-grade parts, with the system’s temperature rating set by its carrier, cooling and enclosure. The AGX Orin Industrial is built for harsher duty at the module itself.

ECC memory detects and corrects single-bit errors, which matters where a corrupted value in a long-running system has consequences; our guide to ECC memory explains how it works. Carriers such as the Forecr DSBOARD‑AGXS are offered with the Industrial module and rated from −40 °C to 85 °C.

On lifecycle, NVIDIA publishes a production lifecycle for every module. The Orin family is listed into the early 2030s, Thor is the newest generation from 2025, and Xavier and TX2 are at end of life. Our answer on how long NVIDIA Jetson is supported sets out the position, and dates move, so confirm against NVIDIA’s lifecycle page before committing a programme.

Specifications only take the decision so far; the rest comes from building these systems. In our own deployments, AGX Orin powers autonomous spot-spraying for agriculture and AI surveillance for transit; Orin NX runs vision-based human-robot safety on the factory floor; and the previous-generation AGX Xavier still ships in counter-drone optical tracking and IP67 waterproof aquaculture systems.

In each case the module followed the application: the camera count, the environment, the power budget and the lifecycle decided it, not the headline figure. The systems below show one example per module family from our range.

We configure, image and test every one with the module, storage and JetPack release fitted, and set the power mode to suit the enclosure, so the module you chose delivers its rating on site.

The right module is the one that disappears into the application: enough compute for the workload, inside the power and thermal budget, with the I/O the system needs and a lifecycle that matches the programme. Over-specifying wastes power, budget and thermal headroom; under-specifying caps the application before it ships.

That sizing is part of every project we take on. Tell us what the system has to do and where it has to run, and we will recommend the module and the system around it, often starting from a carrier board or rugged computer already in our NVIDIA Jetson range.