SDK Installation

This guide covers installing Holoscan SDK development artifacts for NVIDIA Developer Kits (arm64) and x86_64 Linux platforms.

Prerequisites

Before you choose an installation method, check that your system meets the Holoscan SDK 5.0 requirements:

• Operating system: The packaged options use Ubuntu 24.04. You can run the development container on another Linux host supported by the NVIDIA Container Toolkit. “Bare metal” host support is limited to the listed Ubuntu version.
• Host Architectures: The SDK supports x86_64 and arm64. On arm64, the Debian packages (architecture sbsa) and the container cover both arm64 servers with discrete GPUs and NVIDIA Jetson and IGX systems running JetPack 7 or IGX OS 2 and later.
• GPU Architectures: Ampere or newer. See a complete list of tested platforms below.
• GPU mode: Holoscan SDK 5.0 uses a single CUDA 13 software stack for discrete and integrated GPUs. JetPack 6 and earlier are not supported; there is no separate iGPU container or package.
• CUDA: Holoscan SDK 5.0 requires CUDA 13. Use an NVIDIA driver compatible with CUDA 13 (R580 or newer) on the host machine. Container users also need Docker and the NVIDIA Container Toolkit.
• Python: Use Python 3.12 or newer. Holoscan SDK wheels use the CPython 3.12 stable ABI and are intended for Python 3.12, 3.13, and 3.14.
• CMake: Building C++ applications against an installed SDK requires CMake 4.0 or newer. Ubuntu 24.04 ships CMake 3.28, so install a newer version with sudo snap install cmake --classic, pipx install cmake, or the Kitware APT repository. The development container already includes it.

Bare-metal package and wheel users must additionally provide the runtime dependencies identified by their selected components.

Tested platforms

Holoscan SDK 5.0 is tested on the following platforms:

IGX Orin (IGX OS 1.x) and JetPack 6 are not supported.

For hardware-specific configuration, see Additional Setup and Third-Party Hardware Setup. For RDMA, see Enabling RDMA.

Install the SDK

There are multiple ways to install and develop with the Holoscan SDK:

Installation methods

docker pull nvcr.io/nvidia/clara-holoscan/holoscan:v5.0.0-cuda13

Not sure what to choose?

If you want everything ready to go, choose the NGC development container as the safest way to ensure core dependencies are present with the expected versions. It is the simplest way to run the installed examples and get started with developing your own C++ and Python Holoscan applications. The benefit does come at a cost:

• large image size due to the numerous dependencies. If you need a lean runtime image, use one of the alternate installation methods.
• standard inconveniences that exist when using Docker, such as more complex run instructions for proper configuration.

If you are confident in your ability to manage dependencies on your own in your host environment, the Holoscan SDK Debian packages provide all the capabilities needed to use the Holoscan SDK.

If you just want to create new applications with the pre-built Inference, Vision, and Visualization operators, use the Holoscan SDK Python wheels. Note that these may require extra work to set up your host environment with runtime dependencies not available via PyPI today.

For other development cases such as custom Debug builds, targeting an unsupported platform, or experimenting with different configuration options or source patches, please start from the Holoscan SDK open source code.

Camera/V4L2 is outside the initial development-container support contract. Install ConnectX kernel drivers and other host hardware support separately, and validate them on your target system.