NVIDIA Holoscan is an open-source software platform that enables high-performance sensor data processing on edge GPU systems. Originally developed as NVIDIA Clara Holoscan for medical imaging, the platform has since evolved into a domain-agnostic solution suitable for industries such as aerospace, defense, robotics, industrial automation, and more.
Holoscan provides developers and system integrators with a unified framework for deploying AI workloads at the edge with scalability, reliability, and real-time performance.
NVIDIA Holoscan includes optimized libraries for network connectivity, data processing, and AI. It provides examples to create and run low-latency data-streaming applications across industries using C++, Python, or Graph Composer. Using the SDK, developers can build pipelines for sensor data processing that meet latency requirements and scale from the cloud to the data center to the edge.
Holoscan simplifies the development of sensor processing pipelines by offering a modular architecture built around operators (reusable software components with defined inputs and outputs) and applications that orchestrate these operators into a complete workflow. Developers familiar with frameworks like GStreamer or Video4Linux will find Holoscan intuitive.
Holohub includes a wide library of reference operators—many highly relevant for aerospace and defense applications:
Another core benefit is Holoscan’s containerized deployment model, which brings three major advantages:
Holoscan also provides performance optimization tools. For example, the Advanced Networking Operator (available on Holohub) supports GPUDirect RDMA and RoCE, reducing CPU load and latency by up to 90% compared to traditional TCP/UDP protocols, while enabling bandwidth up to 100 Gbps on supported NICs such as the ConnectX-7.
Furthermore, Holoscan integrates with NVIDIA’s performance-optimized libraries, such as TensorRT. Using TensorRT-LLM, you can achieve up to a 70% performance increase when deploying models like LLaMa.
At GTC 2025, EIZO Rugged Solutions showcased a live demo featuring NVIDIA Holoscan running on our Jetson AGX Orin-based Condor AGX-IOX single board computer. Using a USB webcam and NVIDIA’s PeopleNet model, we performed real-time object detection at the edge—drawing bounding boxes around individuals passing our booth.
Our demonstration was powered by the Condor AGX-IOX, a rugged, SOSA®-aligned 3U OpenVPX single board computer designed for extreme environments. Built with the NVIDIA Jetson AGX Orin module, the Condor AGX-IOX offers high-performance AI processing in a compact and durable 3U VPX form factor.
The Condor AGX-IOX was integrated into a custom chassis and backplane with a dedicated rear transition module (RTM) that enabled USB webcam input. With minimal setup, we leveraged the open-source tao_peoplenet application from Holohub to deliver a live demo that performed efficient edge-based object detection.
The diagram below outlines how this system is organized from a high-level view:
As shown above, Holoscan runs entirely on the Condor AGX-IOX. Thanks to containerization, this solution can easily be redeployed to other Condor AGX-IOX units—ideal for field demonstrations or operational deployment across multiple platforms and vehicles.
These instructions were validated on a Condor AGX-IOX using the default board support package provided by EIZO Rugged Solutions. The same process may work on a Jetson AGX Orin Developer Kit.
sudo groupadd docker
sudo usermod -aG docker $USER
newgrp docker
sudo apt install -y cuda-toolkit
sudo nvidia-ctk runtime configure –runtime=docker
sudo systemctl restart docker
sudo apt install -y git
export v4l2_lib=”/usr/lib/aarch64-linux-gnu/libv4l2.so”
sudo mv ${v4l2_lib}.0 ${v4l2_lib}.0.old
sudo ln -s ${v4l2_lib}.0.0.0 ${v4l2_lib}.0
git clone https://github.com/nvidia-holoscan/holohub
cd holohub
git checkout tags/holoscan-sdk-2.9.0
./dev_container build_and_run tao_peoplenet
It may take a few minutes to compile and run the model as TensorRT needs to compile the PeopleNet model to a format that is highly optimized to run for your GPU. But after it has completed, you should see a the webcam display with green bounding boxes drawn around people detected, and red bounding boxes drawn around faces:
There are some issues that may prevent the tao_peoplenet application from running. If you see this error in your terminal window for example:
Traceback (most recent call last):
File “/workspace/holohub/applications/tao_peoplenet/tao_peoplenet.py”, line 370, in <module>
app.run()
RuntimeError: Unsupported pixel format NV12
You may need to make modifications to the file applications/tao_peoplenet/tao_peoplenet.py:
…
source = V4L2VideoCaptureOp(
self,
name=”v4l2_source”,
allocator=pool,
pixel_format=”YUYV”, # Add this line
**v4l2_args,
)
…
Likewise, if the following error is encountered:
ERROR: failed to solve: process “/bin/sh -c pip install -r /tmp/benchmarking_requirements.txt” did not complete successfully: exit code: 1
You may need to make modifications to benchmarks/holoscan_flow_benchmarking/requirements.txt to set pygobject to 3.50.0:
pygobject==3.50.0
…
This is due to the latest version of pygobject relying on girepository-2.0, which is not available for versions of Ubuntu prior to 24.04. There is more information about this here.
EIZO Rugged Solutions offers NVIDIA-based VPX hardware platforms that integrate seamlessly with Holoscan to support AI-powered tasks across embedded edge architectures. Whether you’re building situational awareness applications or deploying multi-sensor edge compute nodes, Holoscan accelerates development and deployment.
Thanks to the portability of containerized applications, developers can build and test on a Jetson AGX Orin Developer Kit and quickly transition to production deployment on a rugged Condor AGX-IOX.
We recommend exploring the following resources to get started:
Looking to integrate NVIDIA’s next-gen GPU architecture? Click below to contact sales for more information.
EIZO Rugged Solutions is a pioneer in providing video, graphics, encoding, and recording solutions for over three decades.
