By NVIDIALearning Objectives
- Understand what the Jetson platform is and why edge AI requires specialized hardware
- Compare the Jetson Orin product lineup and identify appropriate use cases for each
- Evaluate when edge deployment on Jetson is preferable to cloud-based AI inference
What Is Jetson?
NVIDIA Jetson is a family of compact, power-efficient AI computing modules designed for edge deployment — running AI models directly on devices rather than in the cloud. Jetson powers robots, drones, industrial cameras, medical instruments, autonomous vehicles, smart city infrastructure, and any application where AI must run locally.
Edge AI matters because many real-world applications cannot tolerate the latency, cost, or connectivity requirements of cloud inference. A robot navigating a warehouse, a quality inspection camera on a manufacturing line, or a drone performing real-time object detection all need AI processing at the point of action — not a round trip to a data center.
Jetson is used by over 1 million developers and has been deployed in production by thousands of companies. The ecosystem includes pre-trained models, optimization tools (TensorRT), and a full Linux-based software stack (JetPack SDK).
🎯Tip
Start with Jetson: The Jetson Orin Nano Developer Kit ($499) is the entry point — a complete AI computer for prototyping and learning edge AI development.
The Jetson Orin Lineup
| Module | AI Performance | Memory | Best For | Price |
|---|---|---|---|---|
| Jetson Orin Nano | Up to 67 TOPS (INT8) | 8GB LPDDR5 | Learning; prototyping; simple edge AI | $499 (dev kit) |
| Jetson Orin NX | Up to 157 TOPS (INT8) | 8GB or 16GB LPDDR5 | Commercial edge products; mid-range robotics | Module pricing (OEM) |
| Jetson AGX Orin | Up to 275 TOPS (INT8) | 32GB or 64GB LPDDR5 | Advanced robotics; autonomous machines; multi-model workloads | $1,999 (dev kit) |
| Jetson Thor (next-gen) | Up to 800 TOPS | Next-gen memory | Humanoid robots; autonomous vehicles; next-gen edge AI | Expected 2025-2026 |
TOPS = Tera Operations Per Second — a measure of AI inference throughput. For context, running a small language model (under 7 billion parameters, quantized) requires roughly 10-30 TOPS; computer vision models for real-time object detection require 5-20 TOPS.
What Can You Run on Jetson?
- Small language models — Phi-4 (3.8 billion), Gemma 4 E2B and E4B, Nemotron-Mini-4B run efficiently on Orin Nano and above
- Computer vision — real-time object detection (YOLO), image classification, pose estimation, depth estimation
- Speech AI — on-device speech recognition and text-to-speech (NVIDIA Riva)
- Robotics — navigation, manipulation, SLAM (Simultaneous Localization and Mapping) via NVIDIA Isaac ROS
- Multi-model pipelines — on AGX Orin, run vision + language + speech models simultaneously
The JetPack SDK
JetPack is Jetson's software development kit — a complete Linux-based environment including:
- CUDA and cuDNN — GPU-accelerated computing libraries
- TensorRT — inference optimization (quantization, layer fusion, memory planning)
- DeepStream — video analytics SDK for multi-camera, multi-stream AI processing
- Isaac ROS — GPU-accelerated robotics packages for ROS 2
- VPI (Vision Programming Interface) — GPU-accelerated computer vision primitives
Strengths
- Dominant edge AI platform — largest developer ecosystem (1 million+) and broadest model support for edge deployment
- Affordable entry point — $499 Orin Nano dev kit makes edge AI accessible to students and hobbyists
- Full NVIDIA software stack — CUDA, TensorRT, Isaac ROS, DeepStream all optimized for Jetson
- Power efficient — 7-60W power envelope; suitable for battery-powered and thermally constrained devices
- Production-ready modules — same Orin modules used in dev kits ship in commercial products (industrial-grade variants available)
- Strong community — active forums, tutorials, and open-source projects
Limitations & Considerations
- Limited to NVIDIA ecosystem — JetPack and optimized tools only run on Jetson hardware
- Memory constraints — even AGX Orin's 64GB limits the size of models you can run versus datacenter GPUs
- Not a general-purpose computer — optimized for AI inference; not intended as a desktop replacement
- Jetson Thor availability — next-gen module announced but not yet widely available; current Orin lineup may feel aging for cutting-edge use cases
- Learning curve — embedded Linux, cross-compilation, and hardware optimization add complexity beyond cloud development
Key Takeaways
- Jetson is NVIDIA's edge AI platform — compact, power-efficient modules that run AI models locally on devices rather than in the cloud
- The Orin lineup ranges from the $499 Orin Nano (67 TOPS, great for learning) to AGX Orin (275 TOPS, production robotics and autonomous machines)
- Edge AI is essential for applications requiring low latency, offline operation, data privacy, or real-time processing — robotics, manufacturing, drones, medical devices, and autonomous vehicles
- The full NVIDIA software stack (CUDA, TensorRT, Isaac ROS, DeepStream) runs on Jetson, making it the most developer-friendly edge AI platform available
