NVIDIA has officially released JetPack 7.2. This is far more than a routine iterative update; it represents a complete “Platform Reset.” Its core mission is to elevate Jetson from a “development board that can run models” to a “production-grade edge node” capable of deploying multi-agent systems, partitioning, and reproducible builds. As an elite partner of NVIDIA, Seeed Studio is thrilled to announce that our entire Jetson Orin series is now fully compatible with this milestone update.

This release not only brings out-of-the-box support for NemoClaw and revolutionary Agent Skills but also significantly enhances AI inference efficiency and production deployment capabilities for edge devices through full-stack component upgrades and system-level optimizations.

📦 The Orin “Family Bucket” Enters the JetPack 7 Era

For the first time in the 7.x product line, JetPack 7.2 supports the entire Jetson Orin family (AGX Orin / Orin NX / Orin Nano). This means existing Orin devices will now share the same OS, container, and BSP foundation as the future Thor platform. For Seeed developers, this eliminates the cost of cross-version maintenance and grants the Orin series full lifecycle support for Ubuntu 24.04 LTS.

Full-Stack Component Upgrade

JetPack 7.2 extends a unified computing stack based on Ubuntu 24.04 LTS and Linux Kernel 6.8 to the Jetson Orin series, marking a significant leap forward for core components:

Component	JetPack 7.2	JetPack 6.2 (Ref)
Jetson Linux / L4T	39.2	36.4.3
Ubuntu rootfs	24.04 LTS	22.04 LTS
Linux Kernel	6.8	5.15
CUDA	13.2.1	12.6
TensorRT	10.16.2	8.6 / 10.x
DeepStream	8	6.4

This alignment with Data Center CUDA 13 significantly reduces the version gap in cross-platform porting, drastically lowering the cost of migrating x86/Arm server code to the edge.

Architecture Evolution & Installation Changes

JetPack 7.2 aligns Orin with the SBSA (Server Base System Architecture) baseline. After flashing AGX Orin to R39.2, it can run mainstream Arm64 containers (such as upstream vLLM containers) directly without repackaging.

⚠️ Critical Migration Notes:

• Unified ISO Installation: The Orin Nano Developer Kit no longer provides an SD card image. You must use the USB ISO installer for system deployment.
• No apt upgrade: Moving from JetPack 6 to 7 is a major OS and Kernel upgrade. Do not simply run apt upgrade. Please use the OTA package or perform a full flash via ISO/BSP.
• Engine Rebuild Required: Engines built with CUDA 13 / TensorRT 10.16 are incompatible with older versions. Existing models must be rebuilt.

⚡ Performance Unleashed: Super Mode & Memory Efficiency Revolution

JetPack 7.2 features deep bottom layer optimizations designed to unleash the full potential of Jetson hardware. The upcoming reComputer mini J5011 also supports Jetpack 7.2 super mode.

Click here to know more about reComputer Robotics j5011

reComputer Robotics j5011-AGX Orin 32GB Super Mode Explained

For budget-sensitive projects, Super Mode offers a highly cost-effective “sweet spot” configuration. By enabling MAXN_SUPER mode, the 32GB module can approach the performance of the 64GB version:

• AI Performance: Increased from 200 TOPS to 241 TOPS (↑ 20.5%)
• GPU Frequency Unlocked: Increased from 930 MHz to 1300 MHz (↑ 39.8%)
• Cost Advantage: Module cost is approximately 45% lower than the 64GB version, with a power limit raised to 60W.

Edge LLM Inference Test: Memory Usage Plummets by 40%

To intuitive demonstrate performance gains in Generative AI scenarios, we conducted inference tests on the Qwen3.5-27B-Q4_K_M model using an AGX Orin 32GB. Thanks to memory optimizations in TensorRT-Edge-LLM 0.8.0（Powered by CUDA 13 optimizations and the increased compute performance provided by Super Mode）, we achieved stunning results:

Metric	JetPack 6.2	JetPack 7.2	Improvement
Memory Usage	24.6 GB	14.7 GB	📉 ~40% Reduction
Prompt Processing	18.2 tok/s	25.8 tok/s	🚀 41.8% Faster
Token Generation	4.3 tok/s	5.5 tok/s	🚀 27.9% Faster

The reclaimed ~10GB of memory makes multi-service co-deployment possible, allowing even the GR00T N1.7 dual-module VLA to load completely on Orin.

🤖 Agentic AI: From NemoClaw to Device Skills

JetPack 7.2 shifts the edge AI development paradigm from “humans writing code and manually tuning” to “AI agents autonomously controlling hardware.”

NemoClaw One-Click Deployment

Prior to JetPack 7.2, running NemoClaw typically required manual kernel upgrades or complex environment configurations. Now, all dependencies and software stacks are pre-configured. Simply run the following command in your terminal to launch the NemoClaw workflow on your Seeed Jetson device:

curl -fsSL nvidia.com/nemoclaw.sh | bash

Jetson Device Skills: Empowering Agents with Hardware Control

NVIDIA has open-sourced the jetson-device-skills repository, providing a suite of native skills designed specifically for edge devices. These allow AI agents to autonomously read hardware info, auto-optimize, and deploy models with one click. Key skills include:

• jetson-diagnostic: Outputs full device status (model, GPU memory load, temperature, high-usage processes) with one click to quickly identify bottlenecks.
• jetson-memory-audit: Precisely calculates system DRAM and NvMap GPU memory usage, solving memory leak issues during long-term operation.
• jetson-inference-mem-tune: Automatically recommends exclusive memory parameters for mainstream engines like vLLM, SGLang, and llama.cpp.
• jetson-llm-benchmark: Uniformly outputs structured performance metrics, automatically calculating token generation speed and latency for easy model comparison.
• jetson-speculative-decoding: Provides Jetson-exclusive EAGLE-3 speculative decoding configurations, significantly boosting generation speeds on devices with small memory.

Tasks that previously took weeks of manual debugging and optimization can now be completed in days by AI agents invoking these skills.

🛡️ Production-Grade Security: Full Coverage from Boot to Disk

To meet the demands of industrial and automotive sectors for trust chains, tamper resistance, and secure updates, JetPack 7.2 provides a six-layer security architecture:

1. Boot Layer: Secure Boot PKC, burning RSA public key hashes based on fuses to ensure bootloader and kernel must be signed by a private key to start.
2. UEFI Layer: UEFI Secure Boot, verifying UEFI payloads like extlinux.conf and kernel images.
3. TEE Layer: OP-TEE Trusted Execution Environment for secure key storage and running Trusted Applications.
4. Disk Layer: NVLuks / LUKS static data encryption, with encryption keys stored in OTP to prevent physical extraction.
5. Update Layer: Supports atomic A/B partition updates via Mender / RAUC, with automatic rollback on failure.
6. Fleet Layer: Remote image distribution and device management combined with Balena / Peridio.

🏁 Get Started Quickly with Seeed Developer Tools

To help you migrate from JetPack 6.x to the new JetPack 7.2 as quickly as possible, Seeed Studio provides convenient developer tools. You can use the Seeed Developer Tool to simplify the BSP flashing and system migration process. This tool is fully adapted to the latest JetPack 7.2 images, ensuring you can OTA your Seeed Jetson devices to the latest system with one click and immediately experience the powerful features of Agentic AI and production-grade security architecture.

Upgrade your Seeed Jetson device today and explore the infinite possibilities brought by JetPack 7.2!

About Author