Get Started with NVIDIA DeepStream SDK

Overview

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

Release Highlights

Release notes

DeepStream 6.1.1 Highlights:

• Version 6.1.1 adds the following items to release 6.1
• Support for NVIDIA JetPack™ 5.0.2
• An updated compute stack

DeepStream 6.1 Highlights:

• Graph Composer UI new look and feel (extension type icons, extension metadata search, graph auto-layout, better subgraph UI, and other ease-of-use updates)
• (Alpha) Graph Composer Windows 10 support (with remote execution only)
• Support for RealSense stereo cameras
• New post-processing plug-in that makes it easier to use custom models
• Support for NMOS (Networked Media Open Specifications), enabling AI workloads to be easily integrated with IP-based infrastructures
• Integration with NVIDIA® Rivermax® SDK. Now supports reception and transmission for compressed and uncompressed streams
• Support for Unified Communication - X Framework (UCX)/RDMA for data transmission across distributed DeepStream pipelines and nodes (x86/dGPU only)
• New metadata serialization/deserialization plug-ins; allows metadata to be embedded within and extracted from encoded video streams (x86/dGPU only)
• Gst-nvstreammux (new) plug-in: Adds audio support and better tuning control enhancements for video conferencing

Operating System

NVIDIA Jetson™: Ubuntu 20.04

NVIDIA Tesla® GPUs (x86): Ubuntu 20.04

Dependencies

Jetson: JetPack: 5.0.2, NVIDIA CUDA®: 11.4, NVIDIA cuDNN: 8.4, NVIDIA TensorRT™: 8.4.1.5 , NVIDIA Triton™ 22.07, GStreamer 1.16.3

Tesla GPUs (x86): Driver: R515+, CUDA: 11.7 Update 1, cuDNN: 8+, TensorRT: 8.4.1.5, Triton 22.07, 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

To ensure compatibility with supported TensorRT versions , users planning to use models developed with TAO Toolkit (formerly TLT) 3.0-21.08 or earlier MUST re-generate the INT8 calibration cache before using them with DeepStream 6.1.1

Those who are using models and the INT8 calibration cache from previous versions of TensorRT will also need to re-generate the cache.

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 is NOT an issue.