NVIDIA's latest JetPack 7.2 release advances the Jetson software-defined platform with out-of-the-box AI agents, deterministic multi-task execution, greater production flexibility, and over 20% performance improvement on the AGX Orin 32GB. PLink's Jetson-based AI edge computing products now fully support JetPack 7.2.

AI Agents, Out of the Box

Setting up an agent environment on Jetson previously involved numerous dependencies and complex configuration, often taking days from scratch to first run. The most significant change in JetPack 7.2 reduces this process to a single command.

NemoClaw is an open-source agent stack from NVIDIA that adds privacy and security controls to OpenClaw. JetPack 7.2 pre-configures all dependencies, enabling deployment with a single command:

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

The development barrier for agentic physical AI applications, including robotics, industrial automation, visual agents, and edge AI systems, is substantially lowered.

Jetson Agent Skills is another major update in JetPack 7.2. It enables AI agents to go beyond being deployed targets and automatically complete key tasks in the development workflow.

JetPack 7.2 provides three types of Agent Skills:

Jetson Linux Customization Skill: Guides the agent to build a BSP from scratch for custom carrier boards. I/O configuration, clock settings, fan control, and power configuration that previously required weeks of manual work can now be handled automatically by the agent, significantly shortening time-to-market for custom Jetson designs.

Memory Optimization Skill: Automatically adjusts full-stack memory allocation from the bootloader through the kernel to user space, reducing redundant processes and building the most memory-efficient configuration for a given workload. More demanding workloads can run on lower memory configurations, directly reducing total cost of ownership.

Model Benchmarking Skill: Helps identify the most efficient model configuration on the target device, covering model benchmarking, inference optimization, and Jetson diagnostics. For example, when building NemoClaw-based applications, these skills can quickly determine which model performs best for a specific task.

Weeks of manual tuning are now automated by the agent.

20% Performance Gain Without Hardware Changes

For Jetson AGX Orin 32GB users, JetPack 7.2 introduces a new Super Mode.

GPU frequency increases from 930 MHz to 1.3 GHz, and AI performance jumps from 200 TOPS to 241 TOPS, an improvement of over 20%. This means the same hardware delivers near-flagship performance. The 241 TOPS figure is close to the flagship AGX Orin 64GB's 275 TOPS, yet the module cost is approximately 45% lower. The 32GB memory configuration achieves near-64GB performance, and the improvement in memory efficiency translates directly into cost advantages.

A software-only upgrade delivering over 20% AI performance improvement. For decision-makers, this is direct cost optimization; for engineers, it means the same hardware can run larger models and handle more complex tasks.

Jetson AGX Orin module comparison with JetPack 7.2

Deterministic Multi-Task Execution and Greater Production Flexibility

Running multiple AI workloads simultaneously is standard on the Jetson platform. Perception, planning, control, generative AI, and safety monitoring share the same SoC, and latency jitter from resource contention has long been a challenge for mixed-criticality systems.

JetPack 7.2 introduces MIG (Multi-Instance GPU) support on Jetson Thor, partitioning a single GPU into two isolated instances, each with dedicated compute, cache, and memory bandwidth. One instance runs inference while the other handles control, with no mutual interference. Specifically, the integrated Blackwell GPU is divided into two instances: the larger partition (12 SMs, 1,536 CUDA cores) for inference, rendering, and general CUDA workloads; the smaller partition (8 SMs, 1,024 CUDA cores) dedicated to robotics, control, perception, or safety-critical tasks.

Combined with the preemptive RT kernel already available in JetPack 7, MIG creates a more deterministic execution environment for physical AI systems. Humanoid robots, autonomous machines, industrial automation, and medical devices, scenarios sensitive to latency, can achieve reliable coexistence of multiple workloads on a single platform.

For teams moving toward production, system image flexibility also matters. JetPack 7.2 provides official Yocto Project support on Jetson for the first time. Yocto Project is a widely used Linux customization framework for embedded systems, and this support includes verified recipes and reference images.

Key advantages of Yocto Project support:

♦Customizability: Build lean images containing only required services and drivers, reducing memory usage and optimizing system performance, without the need to trim Ubuntu L4T images.

♦Reproducibility: Each build generates identical images, simplifying debugging, testing, and certification processes. This is critical in regulated industries such as medical and industrial.

♦Open ecosystem: Access thousands of packages and community layers for AI frameworks, industrial protocols, and custom middleware.

Additionally, JetPack 7.2 extends the unified compute stack based on Ubuntu 24.04, kernel 6.8, and CUDA 13.0 from Jetson Thor to the Orin series. Both platforms share the same software foundation, enabling a single codebase to deploy seamlessly across Orin and Thor, significantly reducing engineering effort for multi-platform adaptation.

PLink's Jetson-based AI edge computing products now fully support JetPack 7.2. Through software upgrades, the same memory can run more workloads and larger models, lowering the deployment barrier and total cost of ownership for edge AI.

For more information about PLink's AI edge computing products and JetPack 7.2 adaptation, please contact us.