The Jetson lineup, demystified
NVIDIA's Jetson family has been an industrial-AI staple for a decade. The 2026 picture: Xavier is end-of-life-ish for new designs, Orin is the active line, and choices come down to Nano / NX / AGX. Here's how we actually pick between them.Orin Nano (8 GB)
Compute: up to 40 TOPS sparse INT8.Power: 7-15 W envelope.
Memory: 8 GB LPDDR5, shared with GPU.
Where it's the right answer: single-camera vision inference at 1080p, models < 4 GB, latency budget > 30 ms. We run a lot of YOLOv8s + classifier pipelines on these.
Where it isn't: anything memory-hungry. 8 GB shared between OS, application, and model is enough for one model and not enough for two large ones at once.
Orin NX (16 GB)
Compute: up to 100 TOPS sparse INT8.Power: 10-25 W.
Memory: 16 GB LPDDR5.
The sweet spot for most industrial deployments. Enough memory to run a detection model + an anomaly model + a small OCR model concurrently. Power envelope is still passive-cool friendly. Costs about 2x Nano, gets about 2.5x the throughput in our typical workloads.
If you're spec'ing a new vision station today, this is the default unless you have a specific reason otherwise.
Orin AGX (32 / 64 GB)
Compute: up to 275 TOPS sparse INT8.Power: 15-60 W.
Memory: 32 GB or 64 GB LPDDR5.
For multi-camera systems, large models (heavy transformers, pre-trained foundation models running on edge), or applications that need genuine GPU compute alongside inference. We use these for cells with 4+ cameras feeding a single inference engine, and for any deployment that runs a custom CUDA kernel.
Active cooling required. 64 GB SKU is hard to source.
Common gotchas
- Thermal throttling. Even within power envelope, the SoC will throttle if the heatsink is undersized. We see cells with under-spec heatsinks lose ~20 % inference performance in summer.
- eMMC vs SD vs NVMe. SD card wear-out is a 6-12-month problem in 24/7 deployments. eMMC is fine for OS and binaries but small (16 GB on Nano dev kit). NVMe for any deployment that writes images locally — and they all do at some point.
- JetPack version drift. JetPack 6 vs 5 is a real difference (CUDA, TensorRT, OS base). Pinning JetPack version per project is part of our deployment manifest.
- CUDA-only libraries. If the project depends on a library that only ships CUDA backends (a lot of medical / scientific code), Jetson is the only ARM edge platform that runs it. Hailo / Coral can't.
The decision tree
- Single small model, latency > 30 ms is fine, power matters → Orin Nano
- 2-3 models concurrent, latency 10-30 ms, balanced power → Orin NX
- Foundation models, multi-camera, custom CUDA → Orin AGX
- Anything that absolutely needs to be passively cooled in a sealed enclosure → consider Hailo instead
One pattern we no longer try
Putting the Jetson outside the cabinet to "let it breathe". The dust ingress kills the heatsink in six months. Inside the cabinet, properly ducted, with a single 80 mm fan dedicated to the Jetson — that's the deployment that lasts.What's your default Jetson? Curious about NX uptake in 2026 vs sticking with Xavier NX for established cells.
