Getting Started with the NVIDIA Jetson Orin Nano Super: JetPack, CUDA, and Edge AI for Makers

This site's AI-at-the-edge coverage so far has been built around accelerators bolted onto a host board — the Raspberry Pi 5 with the Hailo AI Kit, the Google Coral USB Accelerator, TinyML on an ESP32. The NVIDIA Jetson Orin Nano Super is a different approach entirely: a standalone single-board computer where the GPU and AI acceleration aren't an add-on, they're the reason the board exists. At $249, it's also the cheapest the Jetson line has ever been, which has pulled it into genuine maker-budget territory instead of just robotics-lab territory.

What You're Working With

SpecDetail AI performance67 TOPS sparse / 33 TOPS dense — a software-unlocked increase over the original Orin Nano's 40 TOPS, on the same silicon GPU1,024 CUDA cores, 32 Tensor cores, 1,020MHz CPU6-core Arm Cortex-A78AE @ 1.7GHz RAM8GB PowerUSB-C or barrel connector; full 67 TOPS performance requires the 25W power mode I/OFour USB 3.2 Type-A (10Gbps), two MIPI CSI camera connectors, M.2 2280 and 2230 slots, 40-pin GPIO header pin-compatible with Raspberry Pi Price$249

The "Super" naming refers specifically to a software unlock: existing Orin Nano and Orin NX boards gained a meaningful chunk of this performance through a firmware/software update rather than new silicon — NVIDIA's own figures put the Orin NX 16GB rising from 100 to 157 TOPS and the Orin NX 8GB from 70 to 117 TOPS on the same hardware. If you already own an older Orin board, check whether a software update gets you most of the way to "Super" performance before assuming you need new hardware.

Why Not Just Use a Raspberry Pi + Hailo Kit?

The Hailo AI Kit approach covered elsewhere on this site pairs a general-purpose Pi 5 with a dedicated inference accelerator chip — a good fit when you already have a Pi-based project and want to add object detection without a full redesign. The Jetson takes the opposite approach: the GPU is the primary compute engine, running CUDA and the full NVIDIA software stack rather than a vendor-specific inference-only accelerator API. That matters most if you're working with models or frameworks built around CUDA/TensorRT in the first place (most PyTorch and research computer-vision work), or if you need the model training/fine-tuning flexibility a general CUDA GPU gives you that a fixed-function inference accelerator doesn't.

First-Time Setup

1. Flash the OS image to a microSD card or NVMe drive using NVIDIA's SDK Manager (run from a separate Ubuntu host machine) or the simpler SD Card Image method for a microSD-only setup — the SD card path is the faster route to a first boot if you don't already have an Ubuntu machine handy.
2. The OS is L4T (Linux for Tegra), NVIDIA's own Ubuntu-based distribution built specifically for Jetson hardware — not stock Raspberry Pi OS or generic Ubuntu, even though the board superficially resembles a Pi in form factor and GPIO layout.
3. Install JetPack, NVIDIA's SDK bundle that layers CUDA, cuDNN, TensorRT, and the multimedia/camera APIs on top of L4T — this is the equivalent of ESP-IDF or the Pico SDK in terms of "the thing you actually develop against," and most Jetson tutorials and pre-trained model repos assume a specific JetPack version, so check compatibility before following an older guide.
4. Set the power mode to 25W (MAXN or the equivalent profile) before benchmarking anything — the board can run at lower power profiles that meaningfully cap performance, and a lot of "the Orin Nano Super isn't that fast" complaints trace back to running the default lower-power profile.
5. Run a sample model from NVIDIA's Jetson AI Lab or the JetPack example repos first before building your own pipeline — confirming the camera, CUDA, and TensorRT stack all work together on a known-good model saves a lot of debugging time versus troubleshooting your own code and the platform setup simultaneously.

Practical Notes

• Active cooling is not optional. Running sustained inference workloads at the 25W profile generates real heat; the included or equivalent fan should be running, not treated as an optional accessory for "if it gets warm."
• Power supply quality matters more than it does on a Raspberry Pi. A GPU drawing sustained load under a marginal USB-C supply is a more demanding load profile than typical Pi usage — use a supply rated for the board's actual peak draw, not just "a USB-C charger I had around."
• The 40-pin header is physically GPIO-compatible with Raspberry Pi HATs, but don't assume software compatibility — libraries written against RPi.GPIO or gpiozero won't talk to Jetson's GPIO stack without a Jetson-specific equivalent library.

Closing Thoughts

The Orin Nano Super is the first Jetson that's genuinely competitive with a Pi-plus-accelerator setup on price while offering a meaningfully more capable and more flexible AI compute path once you're past initial setup. The tradeoff is a steeper software stack (L4T and JetPack instead of a familiar Raspberry Pi OS image) and a narrower general-purpose ecosystem — it's a GPU-first board that happens to look like an SBC, not an SBC that happens to have a GPU, and treating it like the latter is the fastest way to end up frustrated with it.