Jetson Orin Specs: Orin Nano, Orin NX and AGX Orin Compared

Jetson Orin specs cover seven production modules, from the Orin Nano 4GB at up to 34 TOPS to the AGX Orin 64GB at 275 TOPS, all built on the same Ampere GPU architecture and the same software stack. That spread, roughly eightfold on one architecture, is why a project can often prototype on one Orin module and ship on another.

NVIDIA Jetson brought GPU-accelerated AI into small, low-power computers, and Jetson Orin is the generation most industrial edge AI systems are built on today. For OEMs and system integrators it is a scalable alternative to an industrial PC with a discrete GPU: one module family covers entry-level vision through to multi-camera robotics, in a fraction of the power.

This reference puts every Orin module’s specifications in one place, with the Super figures from JetPack 6.2, the I/O, the power modes and the Industrial module, and then shows what each module is used for. Figures are NVIDIA’s published values; always confirm against the current module datasheet before fixing a design.

The master table lists the compute, memory and power of each module. Orin Nano and Orin NX are shown in their Super modes with the original figure in brackets.

Storage differs by family. AGX Orin modules carry 64 GB of eMMC 5.1 on the module; Orin NX and Orin Nano have none and boot from an NVMe SSD on the carrier. CPU cache follows the core count, from 1.5 MB of L2 and 4 MB of L3 on the six-core modules to 3 MB of L2 and 6 MB of L3 on the 12-core AGX Orin.

TOPS are sparse INT8, which assumes a pruned network; dense throughput is about half. Our explainer on what TOPS means covers how the figure is counted, and CUDA cores vs Tensor Cores explains which part of the GPU does the deep learning work.

Interfaces often narrow the choice faster than TOPS. These are module capabilities; the carrier board decides which of them reach a connector.

Two points stand out. Orin Nano has no hardware video encoder, so recorders and video gateways usually start at Orin NX; our answer on Orin Nano vs Orin NX covers that difference. And AGX Orin brings 16 camera lanes and 10GbE from the module, which is why multi-camera vehicles and robots usually use it. Our explainers on MIPI CSI-2 and GMSL2 cover the two camera links most Orin systems use.

NVIDIA’s original-mode specification chart, reproduced below, adds the full encode and decode stream counts, the low-speed I/O and the module connectors. Its AI figures predate the Super modes.

JetPack 6.2 added Super modes for Orin Nano and Orin NX, raising GPU and CPU clocks, memory bandwidth and the power ceiling on the same silicon. Existing modules gain them with a software update; nothing changes on the module. Our answer on Jetson Super mode explains the mechanism, and the Orin Nano Super white paper covers the upgrade in depth.

The catch is heat. A Super mode can draw up to 25 W on Orin Nano and 40 W on Orin NX, so a fanless enclosure sized for the original modes may not hold the clocks. Catalog ratings show the trade: Premio rates its fanless JCO-1000-ORN to 55 °C at 15 or 25 W but to 35 °C at 40 W. Decide the power mode with the enclosure, then validate at the hottest ambient.

Every Orin module has a set of NVIDIA power modes, selected with the nvpmodel tool. The mode caps the power and sets the clocks, so it decides the cooling, the power supply and the temperature the system can run at.

System power is always above module power, because the carrier, NVMe, cameras and USB devices draw on top. Worked example: an Orin NX 16GB at 25 W, two PoE cameras at about 7 W each, an NVMe SSD at about 5 W and carrier overhead of a few watts puts the system near 45 W, so a 60 W supply leaves sensible margin. Run the same system at 40 W and the supply, the heatsink and the battery all need resizing.

MAXN modes remove the caps and are best treated as a benchmark setting unless the cooling is designed for them. Our guide to thermal throttling explains what happens when a module runs out of thermal margin, and the Jetson thermal design guide covers fans, heatsinks and conduction-cooled heat spreaders.

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

ECC memory detects and corrects single-bit errors, which matters in long-running systems where a corrupted value has consequences; our guide to ECC memory explains how it works. The Industrial module gives up some peak TOPS for its rating, which is the usual trade on cold or exposed sites. Neousys, for example, rates its Industrial-module NRU-240S-AWP-JAOi outdoor computer from −40 °C to 70 °C in its 35 W mode, and our answer on the industrial temperature range explains what such ratings cover.

TOPS get the attention, but memory is often what rules a module in or out. On Jetson the CPU and GPU share one pool, so the operating system, the application, every model’s weights and the camera buffers all sit in the same memory.

The figures are planning estimates; quantisation, resolution and batch size move them a long way, so measure the real application with tegrastats before fixing a module. Bandwidth matters too: the Super modes raise Orin Nano 8GB from 68 to 102 GB/s, and AGX Orin doubles it again to 204.8 GB/s, which is why camera-heavy pipelines often scale better on the larger modules than the TOPS alone suggest.

Worked example. A four-camera inspection cell runs a detector at 30 frames per second on each stream and records the results. On an Orin Nano 8GB it fits in memory but has no hardware encoder, so recording falls on the CPU. On an Orin NX 8GB the encoder takes the recording, the DLA can run part of the detector and the GPU keeps headroom, so the cell holds its frame rate at 25 W. The NX is the right module, not because of its TOPS, but because of its encoder and memory bandwidth.

For language and vision language models, our explainers on a local LLM and a vision language model cover why memory size sets the ceiling.

Orin is not the only way to run AI at the edge. An x86 industrial PC with a discrete GPU card is the other common route, and each suits different jobs.

Orin’s advantages come from integration. The CPU, GPU, memory and power management sit on one module, so the system is smaller, uses less power and needs fewer parts to fail. The modules are built for edge AI rather than adapted from graphics cards, and they run the same CUDA software as NVIDIA’s data-centre GPUs, so models move from development to deployment with little change. We integrate Orin carrier boards in factory cabinets and custom enclosures with heat spreaders that draw heat away from the module, and every Orin system we ship comes with the carrier vendor’s BSP and a JetPack image.

Where a project needs the most GPU compute per system, a full-size card is the better form factor, and our industrial PCs with GPU cover that route. For the wider comparison, our guide to what NVIDIA Jetson is sets Jetson against NPUs and Raspberry Pi-class boards, and NPU vs GPU for edge AI covers the accelerator choice.

The spec sheet narrows the field; the application decides it. The table below is where each family usually lands in the industrial projects we support.

Orin Nano

The entry-level module is a large step up from the original Jetson Nano, which launched in 2019: NVIDIA quoted the Orin Nano at up to 80 times its predecessor’s AI performance, and the Super modes raise that further. It suits small automation tasks with modest processing needs, installations limited by space or power, and prototyping, where it gives a practical, low-cost way to test whether AI fits a process.

Orin NX

Orin NX keeps the footprint and pin-out of the Orin Nano with up to 157 TOPS in 10 to 40 W. It balances performance and efficiency for medium-scale factory automation, handles vision tasks such as object detection, measurement, tracking and surveillance across several cameras, and is flexible enough to standardize on across a whole manufacturing site.

AGX Orin

AGX Orin is the most capable Orin module, up to 275 TOPS, in a slightly larger form that still fits a compact carrier board or box PC. It suits compute-intensive automation, advanced robotics and cobotics, sophisticated computer vision and large deployments where many systems run in parallel. In our own projects it powers agricultural spot spraying and AI surveillance at stations, while Orin NX runs human-robot safety on the factory floor.

Before Orin, the same jobs ran on the Jetson Xavier generation, the AGX Xavier and Xavier NX, which are now at end of life; our comparison of Jetson Orin vs Xavier covers the step. For a structured way to choose between the Orin modules and Thor, see our Jetson modules comparison, and for the step between the two larger modules, Orin NX vs AGX Orin.

The facts below come from each product’s own specification.

We configure, image and test each system with the module, storage and JetPack release fitted, and set the power mode to suit the enclosure.