AI Performance (TOPS)

The headline number is sparse INT8 throughput. The Orin Nano 8GB hits 67 TOPS in Super mode against roughly 34 TOPS for the 4GB module and well under 1 TOPS-equivalent for the classic Maxwell Nano. Match it to your model: object detection is cheap, transformers are not.

Memory and Bandwidth

CPU and GPU share one pool, so RAM is the hard ceiling. 8GB of 128-bit LPDDR5 at 102 GB/s runs 7B to 8B quantised LLMs and multi-stream vision. 4GB at 51 GB/s is fine for detection and classification but will not hold a useful language model.

Storage and Boot Media

A microSD card is the single biggest performance trap on Jetson. Boot from an NVMe SSD in the M.2 Key M slot instead. Kits that ship with a pre-flashed 128GB or 256GB SSD save you both the money and the flashing session.

Power and Cooling

Orin Nano runs 7W to 25W, and Super mode leans on the top of that range. Active cooling is standard on dev kits; fanless industrial boxes trade peak clocks for silence and reliability. Size your supply for the 25W mode, not the idle draw.

I/O and Camera Interfaces

Two 4-lane MIPI CSI connectors, 4x USB 3.2 Gen2, Gigabit Ethernet, DisplayPort or HDMI, a 40-pin GPIO header and M.2 Key E for Wi-Fi cover most builds. Industrial variants add RS-485, CAN and isolated digital I/O for machinery.

Waveshare packages the genuine NVIDIA Orin Nano 8GB module on the reference carrier and then sells it in a range of bundles that quietly fix the official kit’s omissions. Depending on the SKU you choose, you get an aluminium alloy case, an AC8265 Wi-Fi and Bluetooth M.2 card, and a 256GB NVMe SSD included rather than bought separately. The silicon and the port layout are identical to NVIDIA’s kit, so every JetPack guide and CSI camera driver applies unchanged. Waveshare also tends to have stock when NVIDIA’s own kit is backordered, which alone makes it worth bookmarking.

Yahboom’s SUB kit is the “one box, nothing else to order” option. It pairs the Orin Nano 8GB module with Yahboom’s own compact carrier and throws in a 256GB NVMe SSD, a power supply and an M.2 wireless card, all set up for JetPack 6.2 in Super mode at 67 TOPS. Yahboom also publishes a large free course library covering Ubuntu 22.04, Python, OpenCV and ROS2 on Jetson, which makes this a strong pick for a classroom or for anyone who wants a guided path rather than a pile of forum posts.

The reComputer J3010 is the same cased, cooled, NVMe-equipped design as the J3011 but built around the 4GB Orin Nano. You drop to a 512-core Ampere GPU with 16 Tensor Cores, 51 GB/s of bandwidth and roughly 34 TOPS in Super mode, and in exchange the board sits in a 7W to 20W envelope that is much easier to feed from a battery or to cool passively in a sealed housing. For classic edge vision work such as YOLO detection, licence plate reading, people counting or a multi-camera DeepStream pipeline, it is still plenty.

When the Jetson has to live in a cabinet on a factory floor rather than on your desk, this is the version to buy. The reComputer Industrial J3011 wraps the Orin Nano 8GB in a passively cooled aluminium chassis with no fan to clog or fail, and swaps hobby I/O for the connectors machinery actually uses: two Gigabit Ethernet ports, RS232/RS-422/RS-485, a CAN interface, four digital in/out channels and three USB 3.2 ports, on top of a 128GB NVMe SSD and wide-range DC input with DIN-rail and wall mounting.

The reComputer J1010 is the original 2019 Jetson Nano, boxed properly: a 4GB Maxwell module with 128 CUDA cores and a quad-core Cortex-A57, 16GB of eMMC, JetPack 4.6.1 pre-installed, an aluminium case, Gigabit Ethernet, HDMI, one USB 3.0 and two USB 2.0 ports, an M.2 Key E slot and a Raspberry Pi compatible 40-pin header, all in a 130 x 120 x 50 mm box drawing about 5W. Buy it for one reason only: to run the enormous back catalogue of classic Jetson Nano tutorials and university coursework exactly as written.

This is the only entry that is not a bare board, and that is the point. The ROSMASTER platform mounts an Orin Nano Super 4GB on a steerable chassis and adds a lidar, a depth camera, motor control and a full ROS2 Humble software stack with worked examples for SLAM mapping, autonomous navigation, voice interaction and vision-based following. If your goal is to learn how perception, localisation and planning fit together rather than to benchmark inference, buying the integrated platform saves months of mechanical and wiring work.

Running local LLMs at the edge?

Get the Jetson Orin Nano Super Developer Kit. 67 TOPS and 102 GB/s handle 7B to 8B quantised models.

Want it working today, not tomorrow?

The Seeed reComputer J3011 ships with JetPack 6.2 already built on a 128GB NVMe SSD.

Official kit out of stock?

The Waveshare Orin Nano Super kit uses the same module and carrier, often with the SSD and Wi-Fi card bundled.

Complete beginner or classroom?

The Yahboom 8GB SUB kit includes SSD, PSU and Wi-Fi plus a large free ROS2 and vision course library.

Battery powered or sealed housing?

The 4GB reComputer J3010 runs 7W to 20W and still delivers around 34 TOPS for vision workloads.

Deploying on a machine or in a cabinet?

The fanless reComputer Industrial J3011 adds RS-485, CAN and isolated digital I/O with no moving parts.

Following classic Jetson Nano tutorials?

The reComputer J1010 is the original Nano on JetPack 4.6, cased and ready for legacy coursework.

Building an autonomous robot?

The ROSMASTER Orin Nano kit gives you lidar, depth camera, chassis and a working ROS2 navigation stack.

Is the original Jetson Nano still worth buying in 2026?

Only for a specific reason. The 2019 Jetson Nano is end of life and frozen on JetPack 4.6.x, which means an old CUDA, an old TensorRT and no path to JetPack 5 or 6. Modern frameworks and model formats increasingly will not build against it. That said, it still runs the huge library of classic Jetson Nano tutorials, university lab exercises and older GitHub projects exactly as written, and boards like the reComputer J1010 package it neatly. If you are following legacy material or reproducing an old project, it is fine. For anything new, buy an Orin Nano.

What is the difference between "Jetson Orin Nano" and "Jetson Orin Nano Super"?

“Super” is a software mode, not new silicon. In December 2024 NVIDIA released a JetPack update that raised the clocks and power ceiling on existing Orin Nano modules, taking the 8GB part from 40 TOPS to 67 TOPS and memory bandwidth from 68 to 102 GB/s, and the 4GB part to roughly 34 TOPS. Any Orin Nano developer kit can reach those numbers once you are on JetPack 6.2 or later and the Super power mode is selected, so an older “non-Super” kit is not a lesser product. It does draw more power in that mode, up to 25W, so check your supply.

Do I need an NVMe SSD, or can I just run from a microSD card?

You can boot from microSD, and you will regret it. Model loading, dataset access, Docker images and JetPack itself are all storage bound, and a microSD card turns a 67 TOPS machine into a waiting room. Fit an NVMe SSD in the M.2 Key M slot and boot from it. A 128GB drive is workable, 256GB or more is comfortable once you have a few container images and model weights. Kits like the reComputer J3011 and the Yahboom SUB kit include a pre-flashed SSD, which removes both the purchase and the flashing step.

4GB or 8GB: how much memory do I actually need?

Jetson shares one memory pool between CPU and GPU, so RAM is a hard ceiling rather than a comfort setting. 4GB is enough for classic computer vision: object detection, classification, segmentation, licence plate recognition and modest multi-camera DeepStream pipelines. 8GB is the minimum for anything transformer-shaped, including vision transformers, vision-language models and 7B to 8B quantised LLMs, which typically want 5 to 6GB just for weights. If there is any chance you will want to run a local language model, buy 8GB.

Can a Jetson Orin Nano run local LLMs, and how fast?

Yes, within limits. The 8GB Orin Nano Super handles quantised models up to roughly 8B parameters, and real-world reports from owners running Ollama or a CUDA-enabled llama.cpp build put throughput in the mid-teens to low twenties of tokens per second for 3B to 8B models. That is comfortably fast enough for an offline assistant, a voice interface or an agent loop on a robot, and far too slow for anything you would put in front of many users at once. Expect to build llama.cpp from source with CUDA enabled rather than relying on a generic ARM package.