Get Started With the NVIDIA DeepStream SDK

DeepStream SDK 6.2

Download DeepStream   Forum   Documentation   Try Launchpad


DeepStream is a GStreamer-based SDK for creating vision AI applications with AI for image processing and object detection.

Release Highlights

Release notes

DeepStream 6.2 Highlights:

  • 30+ hardware accelerated plug-ins and extensions to optimize pre/post processing, inference, multi-object tracking, message brokers, and more
  • New nvdsxfer plug-in that enables NVIDIA NVLink? for data transfers across multiple GPUs
  • New REST-APIs that support controle of the DeepStream pipeline on-the-fly
  • Support for new dewarper projections
  • New samples apps: REST-API app, lidar app, industrial inspection app with Basler camera support, and more
  • Updated Graph Composer 2.5 with easy-to-use UI
  • Enterprise support included with NVIDIA AI Enterprise
  • Access to the world’s best performing multi-object trackers: NvDCF, NvDeepSORT, and NvSORT.
  • Support for GigE cameras and lidar sensors
  • Integration with the latest NVIDIA TAO and pretrained models
  • The ability to create custom bindings and scripts to install Python on any DeepStream containers
  • Support for NVIDIA JetPack? 5.1

Operating System

NVIDIA Jetson?: Ubuntu 20.04

NVIDIA Tesla? GPUs (x86): Ubuntu 20.04


Jetson: JetPack: 5.1 , NVIDIA CUDA?: 11.4, NVIDIA cuDNN: 8.6, NVIDIA TensorRT?: , NVIDIA Triton? 23.01, GStreamer 1.16.3

T4 GPUs (x86): Driver: R525+, CUDA: 11.8 , cuDNNs: 8.7+, TensorRT:, Triton 22.09, GStreamer 1.16.3

Note: For JetPack 4.6.1, please use DeepStream 6.0.1. Previous versions of DeepStream can be found here.

Product Advisory

If you’re planning to bring models that use an older version of TensorRT (, make sure you regenerate the INT8 calibration cache before using them with DeepStream 6.2.

You can find details regarding regenerating the cache in the Read Me First section of the documentation. For new DeepStream developers or those not reusing old models, this step can be omitted.

Download DeepStream SDK 6.2

I Agree To the Terms of the NVIDIA DeepStream SDK 6.2 Software License Agreement to Enable Download Links.

DeepStream 5.x applications are fully compatible with DeepStream 6.2. Please read the migration guide for more information.

Python Bindings

The Python bindings source code and pre-built wheels are now available on GitHub.

Introduction to DeepStream SDK

Quick Start Guide

Get step-by-step instructions for building vision AI pipelines using DeepStream and NVIDIA Jetson or discrete GPUs.

Get Started

Introductory DeepStream Webinar

The next version of DeepStream SDK adds a new graph execution runtime (GXF) that allows developers to build applications requiring tight execution control, advanced scheduling and critical thread management

Watch Webinar

Introductory Jetson and Graph Composer Webinar

Learn how NVIDIA DeepStream and Graph Composer make it easier to create vision AI applications for NVIDIA Jetson.

Watch Webinar

Get Started

Find everything you need to start developing your vision AI applications with DeepStream, including documentation, tutorials, and reference applications.

Getting Started with C/C++

Get Started

Getting Started with Python

Learn how the latest features of DeepStream are making it easier than ever to achieve real-time performance, even for complex video AI applications.

Get Started Python Application
GitHub Repository
Compile and Install
Python Bindings
Python Sample Applications

Getting Started with Graph Composer

Learn how NVIDIA DeepStream and Graph Composer make it easier than ever to create vision AI applications for NVIDIA Jetson.

Get Started

Additional Resources

Ethical AI
NVIDIA 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. Also, 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.