DeepStream SDK | NVIDIA NGC

NVIDIA

DeepStream SDK

Collection

NVIDIA

DeepStream SDK

NVIDIA DeepStream SDK enables developers to build accelerated pipelines for a wide range of use-cases such as IVA, retail, industrial inspection, and many more with minimal development effort.

NVIDIA DeepStream SDK

DeepStream is NVIDIA's GPU-accelerated toolkit for building real-time video, image, and sensor analytics pipelines.

Learn about DeepStream Skills. Reusable building blocks for perception, tracking, and reasoning. Paired with DeepStream coding agents, Skills take vision AI development from weeks to hours.

DeepStream 9.1 is now GitHub-first

Starting with 9.1, the DeepStream monorepo is the central repository for the SDK, its tools, and its skills. This introduces multiple changes to how you get DeepStream.

DeepStream now lives on GitHub.
Installers. Available on GitHub Releases, no longer on NGC.
Containers. Consolidated on NGC at nvidia/deepstream.
Legacy Jetson containers. DeepStream 9.0 and older remain at nvidia/deepstream-l4t.

Full documentation, tools, and API reference are on the DeepStream monorepo on GitHub and in the DeepStream Developer Guide.

Available containers

Container	Platforms	Inference backend	Best for
`deepstream:9.1-triton-multiarch`	x86 + Jetson	Triton (PyTorch, ONNX, TensorRT)	Development and deployment
`deepstream:9.1-samples-multiarch`	x86 + Jetson	TensorRT	Deployment
`deepstream:9.1-triton-sbsa-dgx-spark`	ARM SBSA, DGX Spark	Triton	Deployment

Dockerfiles for the above containers are available in the DeepStream monorepo for easy customization.

Installers and libraries

Artifact	Release 9.1 or newer	Release 9.0 or older
DeepStream installers	GitHub Releases	NGC resource page
DeepStream Libraries	GitHub Releases	NGC resource page

DeepStream SDK 9.1 Release

For a full list of new features, changes, and known limitations, refer to the DeepStream 9.1 Release Notes.

Key New Features and Enhancements available in DeepStream 9.1:

Category	Details
Release Highlights	New agentic skills help fast-track vision AI development. A multi-camera 3D tracking (MV3DT) skill enables accurate, end-to-end tracking across multiple cameras. The AutoMagicCalib (AMC) skill automates camera calibration, saving time and reducing errors. NVIDIA JetPack™ 7.2 support accelerates vision AI on the NVIDIA Jetson™ edge AI platform (Orin™, Thor™). Open-source access in a unified GitHub repo simplifies adoption and maintenance.
Enhancements	Please refer to the release notes section of the documentation

Category

Details

Release Highlights

New agentic skills help fast-track vision AI development.
A multi-camera 3D tracking (MV3DT) skill enables accurate, end-to-end tracking across multiple cameras.
The AutoMagicCalib (AMC) skill automates camera calibration, saving time and reducing errors.
NVIDIA JetPack™ 7.2 support accelerates vision AI on the NVIDIA Jetson™ edge AI platform (Orin™, Thor™).
Open-source access in a unified GitHub repo simplifies adoption and maintenance.

Enhancements

Please refer to the release notes section of the documentation

Supported Hardware and Software Dependencies

Embedded Platforms

Jetson Series	Modules	JetPack	Latest DeepStream Supported
Thor	Jetson AGX Thor	7.2	9.1
Orin	Jetson AGX Orin, Jetson Orin NX, Jetson Orin Nano IGX (including dGPU mode)	7.2	9.1
Xavier	Jetson AGX Xavier, Jetson Xavier NX	5.1.3	6.3 (legacy)
TX2	Jetson TX2 NX, Jetson TX2	4.6.x	6.0 (legacy)
Nano	Jetson Nano	4.6.x	6.0 (legacy)

Enterprise Platforms

Enterprise GPU Architecture	GPUs
Blackwell	B100, B200, RTX PRO Server and Workstation Edition, GB200, RTX PRO 4500
Ada Lovelace	L4, L40, L40S
Ampere	A2, A10, A16, A30, A40, A100, RTX A6000
Hopper	H100, H200
Turing	T4

Software Dependencies

Software Component	x86	ARM SBSA / Jetson (via JetPack)
Operating System	Ubuntu LTS 24.04	Ubuntu LTS 24.04
GStreamer	1.24.2	1.24.2
Rivermax	1.80.24	1.70.32
DLFW (Triton)	26.03	26.04
TensorRT	10.16.0.72	10.16.1.7
CUDA	13.2	13.2
cuDNN	9.20.0.48	9.20.0.46
GPU Driver (RM)	595+	595+
CV-CUDA (DeepStream Libraries)	0.16.0	N/A - Jetson 0.16.0 - ARM SBSA
NvImageCodec (DeepStream Libraries)	0.8.0.22	N/A - Jetson 0.8.0.22 - ARM SBSA
PyNvVideoCodec (DeepStream Libraries)	2.1	N/A - Jetson 2.1 - ARM SBSA

Note:

If you are looking for older versions of DeepStream, please refer to the x86 or Jetson archive. Archived documentation is available here.

Technical Support

DeepStream Support is available via:

Method	Available to
Forums	All developers
Direct Support	NVIDIA AI Enterprise License holders

Known Limitations

There are known bugs and limitations in the SDK. To learn more about those, refer to the release notes.

License

The following licenses apply to the DeepStream SDK assets:

Asset	Applicable EULA	Notes
SDK	DeepStream SDK EULA	A copy of the license is available on the following folder of the SDK: /opt/nvidia/deepstream/deepstream-9.1/LicenseAgreement.pdf
Containers	DeepStream NGC License	License grants redistribution rights allowing developers to build applications on top of the DeepStream containers
Development Containers	DeepStream NGC Development License	A development-only license. Does not allow redistribution of the container
TAO Models	NVIDIA AI Product License	All TAO pre-trained models included in the DeepStream SDK are covered by the NVIDIA AI Product License.
DeepStream on GitHub	DeepStream OSS License	Applies to the DeepStream source, tools, and skills hosted on GitHub.

NOTE: By pulling, downloading, or using the DeepStream SDK, you accept the terms and conditions of the EULA licenses listed above.

Ethical AI

NVIDIA’s platforms and application frameworks enable developers to build a wide array of AI applications. Consider potential algorithmic bias when choosing or creating the models being deployed. Work with the model’s developer to ensure that it meets the requirements for the relevant industry and use case; that the necessary instruction and documentation are provided to understand error rates, confidence intervals, and results; and that the model is being used under the conditions and in the manner intended.