NVIDIA Jetson AGX Orin: The 275 TOPS Edge AI Module for Robotics and Physical AI
The Jetson AGX Orin packs a 2048-core Ampere GPU, a 12-core Arm Cortex-A78AE CPU, and 64GB of LPDDR5 memory into a production-ready system-on-module. With configurable power from 15W to 60W and the full NVIDIA AI software stack, it is the platform for autonomous machines, robotics, and agentic AI at the edge.
Affiliate Disclosure: This article contains affiliate links. MiniPCDeals.net participates in the Amazon Associates program and may earn a commission on qualifying purchases at no extra cost to you. This does not affect the information presented here, which is based on manufacturer-published specifications.
The NVIDIA Jetson AGX Orin is the flagship system-on-module in the Jetson Orin family, delivering up to 275 TOPS of AI performance with power configurable between 15W and 60W. It combines a 2048-core Ampere GPU, a 12-core Arm Cortex-A78AE CPU, and 64GB of 256-bit LPDDR5 memory. With JetPack 7.2 and NemoClaw support, it is agentic-ready — designed for robotics, autonomous machines, industrial inspection, and physical AI applications at the edge.
What Is the Jetson AGX Orin?
The Jetson AGX Orin is the most powerful member of NVIDIA’s Jetson Orin family of embedded AI computers. It is a system-on-module — not a consumer device — designed to be integrated into robots, autonomous vehicles, industrial inspection systems, and medical devices.
Unlike a desktop GPU or a mini PC, the AGX Orin is a building block. It ships as a module that OEMs and developers mount onto custom carrier boards, giving them full control over I/O, enclosure, and power delivery. This is what makes it suitable for production deployments at scale.
The Orin family includes seven modules sharing the same architecture. The AGX Orin sits at the top of that range, offering more CUDA cores, more memory, and a wider power envelope than the Orin NX or Orin Nano series. It is available in 64GB, 32GB, and Industrial variants.
The Jetson AGX Orin Developer Kit is a complete desktop-sized board with ports, cooling, and power supply — ideal for prototyping. The module is the bare SoM for integration into production hardware. Both share the same SoC architecture, so code developed on the kit runs unchanged on the module.
Full Specifications
| AI Performance | 275 TOPS (sparse INT8) |
|---|---|
| GPU | NVIDIA Ampere architecture · 2048 CUDA cores · 64 Tensor Cores · max 1.3 GHz |
| CPU | 12-core Arm Cortex-A78AE v8.2 64-bit · 3MB L2 + 6MB L3 · max 2.2 GHz |
| Memory | 64GB 256-bit LPDDR5 · 205 GB/s bandwidth |
| Storage | 64GB eMMC 5.1 |
| Deep Learning Accelerators | 2× NVDLA v2.0 |
| Vision Accelerator | PVA v2.0 |
| Power | Configurable 15W–60W |
| Networking | Up to 10 GbE (via carrier board) |
| Camera | Up to 16 lanes MIPI CSI-2 |
| Video | Encode: 2× 4K60 or 16× 1080p30 · Decode: 8K30 or 22× 1080p30 · H.265, H.264, VP9, AV1 |
| I/O (via carrier) | PCIe Gen4 · USB 3.2 · DP 1.4 · CAN · SPI · I²C · GPIO |
| Software | JetPack SDK · CUDA · TensorRT · Isaac · DeepStream · Riva |
Architecture Explained
The Jetson AGX Orin is built around a heterogeneous compute architecture: the Ampere GPU handles parallel AI workloads, the Arm CPU manages control flow and system tasks, the NVDLA accelerators run inference at low power, and the PVA handles vision processing — all sharing the same memory pool.
This design lets multiple AI pipelines run concurrently. A single AGX Orin can process video streams for object detection, run a language model for natural language understanding, and handle sensor fusion for 3D perception — without switching between discrete accelerators.
The memory subsystem is equally important. With 205 GB/s of bandwidth and 64GB of capacity, the module can hold large models and high-resolution sensor data simultaneously. This is a significant step up from the 1024-core, 16GB configuration of the Orin NX, and it is what makes the AGX Orin suitable for multimodal AI at the edge.
At 205 GB/s, the AGX Orin’s LPDDR5 bandwidth is impressive for an embedded module, but it is far below the 273 GB/s of the DGX Spark or the 1,792 GB/s of an RTX 5090. The AGX Orin is not designed to run frontier 200B-parameter models locally. Its strength is real-time inference on multimodal models — vision, speech, sensor fusion — within a 15–60W power envelope.
Configurable Power Modes
The AGX Orin supports power modes from 15W to 60W, allowing developers to trade performance for efficiency depending on the deployment. A battery-powered drone can run at 15W; a wall-powered industrial robot can use the full 60W.
This configurability is managed through NVIDIA’s power mode system. Each mode defines CPU and GPU clock limits, the number of active cores, and the power budget. Developers can switch modes at runtime through software — useful for robots that need peak performance during navigation but lower power when idle.
- 15W mode: battery-powered edge devices, drones, mobile robots
- 30W mode: balanced performance for most robotics applications
- 60W mode: maximum throughput for autonomous vehicles and industrial inspection
NVIDIA Jetson AGX Orin — 275 TOPS · 2048 CUDA cores · 64GB LPDDR5
Available on Amazon. Production-ready system-on-module for robotics and physical AI.
Position in the Orin Family
The AGX Orin sits at the top of the Orin family. Below it are the Orin NX (up to 100 TOPS, 10–25W) and the Orin Nano (up to 67 TOPS, 7–25W). All three share the same architecture, so software developed for one runs on the others.
| Module | AI TOPS | GPU Cores | Memory | Power |
|---|---|---|---|---|
| Jetson AGX Orin 64GB | 275 | 2048 | 64GB LPDDR5 | 15–60W |
| Jetson AGX Orin 32GB | 241* | 2048 | 32GB LPDDR5 | 15–60W |
| Jetson Orin NX 16GB | 100 | 1024 | 16GB LPDDR5 | 10–25W |
| Jetson Orin Nano 8GB | 67 | 512 | 8GB LPDDR5 | 7–25W |
* Jetson AGX Orin 32GB received a performance boost to 241 TOPS, up 20% over its original specification, via a JetPack update.
The choice between modules comes down to workload. The AGX Orin is for systems that need maximum AI throughput and can accommodate a larger power budget — autonomous vehicles, industrial robots, medical imaging devices. The Orin NX and Nano target smaller, lower-power devices where 100 TOPS or less is sufficient.
Use Cases
The Jetson AGX Orin is designed for edge deployments where cloud connectivity is unreliable, latency matters, or data cannot leave the device. Its typical applications are robotics, autonomous systems, industrial inspection, and medical devices.
Robotics and Autonomous Machines
The AGX Orin powers mobile robots, collaborative robotic arms, and autonomous vehicles. It handles perception, planning, and control in a single module, running frameworks like NVIDIA Isaac for robotics development and deployment.
Industrial Inspection and Smart Cities
With support for up to 16 MIPI CSI-2 camera lanes and 10 GbE networking, the module can process multiple high-resolution video streams for defect detection, traffic monitoring, and public safety applications. NVIDIA Metropolis provides the video analytics framework.
Medical Devices and Healthcare
The AGX Orin is used in medical imaging systems, surgical robots, and patient monitoring devices where on-device inference is required for privacy and latency reasons.
Physical AI and Agentic Systems
With JetPack 7.2 and NemoClaw, the AGX Orin can run agentic AI workflows at the edge — AI agents that perceive their environment, make decisions, and act autonomously. This is the emerging category NVIDIA calls physical AI.
A published study used a Jetson AGX Orin module for real-time predictive maintenance of collaborative robotic arms, running a physics-informed digital twin model for hardware-in-the-loop validation over 50 autonomous cycles. The AGX Orin handled the inference locally, without cloud connectivity.
Software Stack and NemoClaw
The AGX Orin runs the complete NVIDIA AI software stack via JetPack SDK. The recent JetPack 7.2 release adds Yocto project support, CUDA 13 on Orin, and NemoClaw support — making Jetson agentic-ready for production deployments.
JetPack 7.2 brings several upgrades relevant to AGX Orin developers. Yocto-based OS support gives industrial customers a leaner, more customizable Linux foundation — important for memory-constrained deployments. CUDA 13 brings the latest compute stack to existing Orin devices.
NemoClaw, NVIDIA’s agentic AI framework, extends Jetson into the physical world. It provides security guardrails for autonomous agents, enabling developers to deploy AI that can read files, call APIs, and execute multi-step workflows — all locally, on the edge.
The full software stack includes:
- JetPack SDK — OS, drivers, CUDA, TensorRT, cuDNN
- NVIDIA Isaac — robotics development and simulation
- DeepStream — vision AI and video analytics
- Riva — conversational AI and speech
- NemoClaw — agentic AI framework with security guardrails
- NGC catalog — pretrained models and containers
Frequently Asked Questions
Sources & Notes
Specifications sourced from NVIDIA’s official Jetson AGX Orin product pages and technical documentation (NVIDIA Jetson AGX Orin Developer Kit data sheet, NVIDIA Jetson Orin module comparison). JetPack 7.2 and NemoClaw information from NVIDIA developer blogs (COMPUTEX 2026 announcements). Real-world deployment example from published research on predictive maintenance with collaborative robotic arms (Springer, 2026). This content is informational and based on manufacturer- and community-published data.