Is this page helpful?

DeepStream SDK 9.0 for NVIDIA dGPU/X86 and Jetson#

DeepStream Release Notes#

1.0 About this Release

1.1 What’s New

1.1.1 DS 9.0

1.1.2 DS 8.0 (Previous Release)

1.2 Differences since Deepstream 6.1 and Above

2.0 Limitations

3.0 Notes

3.1 Deploying Applications in a Docker Container

About this Release#

These release notes are for the NVIDIA® DeepStream SDK for NVIDIA® Turing ®, NVIDIA® Ampere®, NVIDIA® Hopper®, NVIDIA® Ada Lovelace®, NVIDIA® Blackwell®, NVIDIA® RTX pro 4500® and NVIDIA® Jetson Thor ™.

What’s New#

The following new features are supported in this DeepStream SDK release:

DS 9.0#

  • Support for RTX pro 4500 GPUs

  • Support for Triton 26.01 on x86 and Triton 25.08 on Jetson

  • Jetson package based on JP 7.1 (r38.4 BSP)

  • Added support for Opentelemetry and Prometheous metric exporter

  • Accuracy improvements for Reference Apps - MV3DT, MaskTracker, Pose Estimation

  • REST API support for Health checkpoints and monitoring metrics

  • Added blur-mask to NvOSD component

  • Enhancements in Tracker - Filter objects based on class ID

  • Support of OBB (Oriented Bounding Boxes)

  • DeepStream Coding Agent support

  • VLM support (Cosmos Reason 2) via nvvllmvlm plugin

  • OpenSourced components

    • nvdsmetainsert

    • nvdsmetaextract

    • nvimageenc

  • Added Pyservice maker support for Smart-Recording

DS 8.0 (Previous Release)#

  • Support for Blackwell and Jetson Thor

  • Jetson package based on JP 7.1 (r38.2 BSP)

  • New applications using pyservicemaker: Action recognition app, Smart record app, DeepStream test5 app using Flow API, Kafka test app using pipeline APIs

  • Added MaskTracker, MediaExtractor, Multi-View 3D Tracking (MV3DT), pose estimation to Single and Multi-View 3D Tracking

  • Support for Dynamic stream handling in demuxer

  • nvvideoconvert to support UYVY (8-bit YCbCr-4:2:2) on x86/dGPU

  • OpenSourced components - nvll_osd library, smart record library “gst-nvdssr”, nvimageenc, nvimagedec, nvdsudpsrc, nvdsudpsink, nvdsanalytics gst-plugins

  • REST API support for nvdsanalytics and nvtracker plugin

  • Added nv3dsink support on x86

  • Support for Inference Builder open-source tool to create inference microservice across multiple AI frameworks

  • Added support for high quality pixel formats in nvvideotestsrc

  • Support for TAO 6.0 models

Differences since Deepstream 6.1 and Above#

gstreamer1.0-libav, libav, OSS encoder,decoder plugins (x264/x265) and audioparsers packages are removed in DeepStream dockers from DeepStream 6.1 onwards. You may install these packages based on your requirement (gstreamer1.0-plugins-good/ gstreamer1.0-plugins-bad/ gstreamer1.0-plugins-ugly). While running DeepStream applications inside dockers, you may see the following warnings:

WARNING from src_elem: No decoder available for type ‘audio/mpeg, mpegversion=(int)4, framed=(boolean)true, stream-format=(string)raw, level=(string)2, base-profile=(string)lc, profile=(string)lc, codec_data=(buffer)119056e500, rate=(int)48000, channels=(int)2’.

Debug info: gsturidecodebin.c(920): unknown_type_cb ():

To avoid such warnings, install gstreamer1.0-libav and gstreamer1.0-plugins-good inside docker.

Specifically, for deepstream-nmos, deepstream-avsync-app and python based deepstream-imagedata-multistream app you would need to install gstreamer1.0-libav and gstreamer1.0-plugins-good.

Gst-nveglglessink plugin is deprecated. Use Gst-nv3dsink plugin for Jetson instead.

Limitations#

This section provides details about issues discovered during development and QA but not resolved in this release.

  • DeepStream on Jetson is based on L4T BSP version r38.4. Refer to “Known Issues” section in Jetson release notes.

  • With V4L2 codecs only MAX 1024 (decode + encode) instances are provided. The maximum number of instances can be increased by making changes in open-source code.

  • detected-min-w and detected-min-h must be set to values larger than 32 in the primary inference configuration file (config_infer_primary.txt) for gst-dsexample on Jetson.

  • The Kafka protocol adapter sometimes does not automatically reconnect when the Kafka Broker to which it is connected goes down and comes back up. This requires the application to restart.

  • If the nvds log file ds.log has been deleted, to restart logging you must delete the file /run/rsyslogd.pid within the container before reenabling logging by running the setup_nvds_logger.sh script. This is described in the “nvds_logger: Logging Framework” sub-section in the “Gst-nvmsgbroker” section.

  • Running a DeepStream application over SSH (via putty) with X11 forwarding does not work.

  • DeepStream currently expects model network width to be a multiple of 4 and network height to be a multiple of 2.

  • Triton Inference Server implementation in DeepStream currently supports a single GPU. The models need to be configured to use a single GPU.

  • For some models output in DeepStream is not exactly same as observed in TAO Toolkit. This is due to input scaling algorithm differences.

  • Dynamic resolution change support is Alpha quality.

  • On the fly Model update only supports the same type of Model with same Network parameters.

  • Rivermax SDK is not part of DeepStream. So, the following warning is observed (gst-plugin-scanner:33257):

    GStreamer-WARNING **: 11:38:46.882: Failed to load plugin ‘/usr/lib/x86_64-linux-gnu/gstreamer-1.0/deepstream/libnvdsgst_udp.so’: librivermax.so.0: cannot open shared object file: No such file or directory

    You can ignore this warning safely.

  • When using Composer WebSocket streaming, sometimes errors like “Error while sending buffer: invalid state” is seen, or the window becomes unresponsive. Refreshing the browser page might fix it.

  • Composer WebRTC Streaming is supported only on RTX GPUs.

  • On jetson, when the screen is idle, fps is lowered for DeepStream applications. This behavior is by design to save power. However, if user does not want screen idle then refer to the FAQ for WAR.

  • RDMA functionality is only supported on x86 and only in x86 Triton docker for now.

  • You cannot build the DeepStream out of the box on Jetson dockers except its Triton variant.

  • There can be performance drop from TensorRT to Triton for some models (5 to 15%). In such cases, user may want to use nvinfer plugin instead nvinferserver plugin.

  • NVRM: XID errors seen sometimes when running 200+ streams on Ampere, Hopper and ADA.

  • NVRM: XID errors seen on some setups with gst-dsexample and transfer learning sample apps.

  • Sometimes during deepstream-testsr app execution, assertion ” GStreamer-CRITICAL **: 12:55:35.006: gst_pad_link_full: assertion ‘GST_IS_PAD sinkpad)’ failed” is seen which can be safely ignored.

  • For some of the models during engine file generation, error “[TRT]: 3: [builder.cpp::~Builder::307] Error Code 3: API Usage Error“ observed from TensorRT, but has no impact on functionality and can be safely ignored.

  • deepstream-server app is not supported with new nvstreammux plugin.

  • TAO point-pillar model works only in FP32 mode.

  • REST API support for few components (decoder, preprocessor, nvinfer along with stream addition deletion support) with limited configuration options. However, you can extend the functionality with the steps mentioned in SDK documentation.

  • While adding and removing streams continuously using REST API at frequent intervals, memory surge is observed. This is due to new libc behaviour on Ubuntu 24. User should export below system environment variables before running application which uses REST API.

    • export MALLOC_ARENA_MAX=1

    • export MALLOC_MMAP_MAX_=0

    • export MALLOC_MMAP_THRESHOLD_=131072

    • export MALLOC_TRIM_THRESHOLD_=131072

  • With Basler camera, on Jetson, only images with width of multiple of 4 supported.

  • In some cases, performance with Python sample apps may be lower than C version.

  • while running deepstream-opencv-test app, warning “gst_caps_features_set_parent_refcount: assertion ‘refcount == NULL’ failed” observed. No impact on functionality & can be safely ignored.

  • Observing below errors for Jetson dockers (but no impact on functionality)

    • While decoding: /bin/dash: 1: lsmod: not found and /bin/dash: 1: modprobe: not found.

  • Performance drop for some Models seen with nvinferserver in gRPC mode when run on ARM SBSA w.r.t to x86.

  • Minor performance drop observed compared to the DS 6.4 release when using the NvDCF performance configuration. To improve performance in this case, set the environment variable NVDS_DISABLE_CUDADEV_BLOCKINGSYNC=1.

  • For Azure, messages sent not matching with messages received on server side.

  • Performance in WSL is not at par with Ubuntu system. There is a known throughput issue while running multiple decode instances in WSL. So you may observe lower FPS compared to ubuntu baseline

  • Image encode not supported in WSL.

  • In WSL2, black screen with log “”MESA: error: Failed to attach to x11 shm” is observed with pipelines with nveglglessink or other display sinks. Use filesink instead in such usecases.

  • Numpy 2.x is not supported by PyDS.

  • While running ~200+ streams simultaneously kernel crash may be observed randomly which may stall the application for some time.

  • Inference builder based deepstream app samples hang on B200 platform when the batch-size exceeds 16.

Notes#

  • DeepStream Python bindings is deprecated. Usage of pyservicemaker instead of python bindings is recommended.

  • GFX/Graph Composer support is deprecated since DS 8.0.

  • NVIDIA® DeepStream SDK 9.0 supports TAO model-based applications (https://developer.nvidia.com/tao-toolkit). For more details, see NVIDIA-AI-IOT/deepstream_tao_apps (branch: release/tao_ds9.0ga).

  • On vGPU, only CUDA device memory NVBUF_MEM_CUDA_DEVICE supported.

  • Usecase with HW encoder + RTSP in long duration may face issue, work around is to restart application.

  • For DGX Spark, as VIC is not supported, need to set tiler compute-hw to GPU (compute-hw=1) Example: In ds_python app - deepstream_test_3.py, set compute_hw=1 for tiler property.

  • For DGX Spark, WARNING: Detected NVIDIA GB10 GPU, which is not yet supported in this version of the container seen while launching container. This is harmless and can be ignored as it is does not impact the functionality.

  • From DeepStream 8.0 onwards, the following models are removed. Facedetect, FacedetectIR, PeopleSegnet, bodyposenet, gesturenet,emotionnet, hearratenet, gazenet, facial landmark estimation models and its corresponding applications. Yolo OSS, SSD, DSSD, Yolov3, Yolov4, Yolov4-tiny, Fasterrcnn, Densenet models and it’s corresponding applications.

  • DeepStream Audio support, ASR, TTS plugin not supported. User to use Nvidia RIVA speech sdk.

  • DeepStream 8.0 onwards removed support for tensorflow, uff and caffe models.

    The Tensorflow Backend has been deprecated starting in 25.03. The last release of Triton Inference Server with the Tensorflow Backend is 25.02. To continue using the Tensorflow backend starting with version 25.03 and onward, users must compile the Tensorflow backend themselves from the source code, since pre-built versions will no longer be available.

    For more details refer: https://docs.nvidia.com/deeplearning/triton-inference-server/release-notes/rel-25-03.html#rel-25-03

  • Removed int8 calibration support for previous DeepStream releases supported TAO models. Lower perf with default model will be observed, as we are moving to FP16 mode from INT8.

  • nvv4l2decoder does not support JPEG decode for Thor OpenRM.

  • DLA is not supported on Jetson Thor

  • For Protocol adapters - Password through config file will be deprecated from the next release.

  • Some of the open-source libraries related to Codecs have been removed so user might see some warnings as below, which can be ignored safely:

For example: (gst-plugin-scanner:1433): GStreamer-WARNING **: 18:05:56.454: Failed to load plugin ‘/usr/lib/aarch64-linux-gnu/gstreamer-1.0/libgstfaad.so’: libfaad.so.2: cannot open shared object file: No such file or directory /bin/bash: line 1: lsmod: command not found /bin/bash: line 1: modprobe: command not found

Note

  • OpenCV Deprecation: OpenCV is deprecated by default. However, you can enable OpenCV support in plugins such as nvinfer (nvdsinfer) and dsexample (gst-dsexample) by setting WITH_OPENCV=1 in the Makefile of these components. Refer to the component README for detailed instructions.

  • Docker Note: If your application requires OpenCV and you are using Docker, ensure that the libopencv-dev package is installed inside the Docker container.

Known Issues#

  • During ultra-long duration runs (60+hr) involving live streams with NvDCF accuracy tracking config and software encoder on specific Ampere SKUs, an intermittent pipeline stability issue may occur involving a g_mutex_clear() error.

  • On Jetson, when using CUDA Unified memory with Deesptream reference app (example: dsexample), you may get the pipeline failure with a NvBufSurfaceUnMapNvmmBufferImpl: Wrong memType(3) error. It is recommended to use CUDA Device/Host memory instead.

  • On Jetson, using the deepstream-redaction-app Python app, you may get the pipeline failure with nvbufsurface: mapping of memory type (2) not supported error when accessing surface data.

  • On DGX Spark, error “Cuda failure: status=1 in createTexture” is observed when running appsrc-cuda-test app with nvinferserver plugin.

Deploying Applications in a Docker Container#

Refer Deepstream docker section for detailed information of using DeepStream SDK inside dockers.

Notice

THE INFORMATION IN THIS DOCUMENT AND ALL OTHER INFORMATION CONTAINED IN NVIDIA DOCUMENTATION REFERENCED IN THIS DOCUMENT IS PROVIDED “AS IS.” NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE INFORMATION FOR THE PRODUCT, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE. Notwithstanding any damages that customer might incur for any reason whatsoever, NVIDIA’s aggregate and cumulative liability towards customer for the product described in this document shall be limited in accordance with the NVIDIA terms and conditions of sale for the product. THE NVIDIA PRODUCT DESCRIBED IN THIS DOCUMENT IS NOT FAULT TOLERANT AND IS NOT DESIGNED, MANUFACTURED OR INTENDED FOR USE IN CONNECTION WITH THE DESIGN, CONSTRUCTION, MAINTENANCE, AND/OR OPERATION OF ANY SYSTEM WHERE THE USE OR A FAILURE OF SUCH SYSTEM COULD RESULT IN A SITUATION THAT THREATENS THE SAFETY OF HUMAN LIFE OR SEVERE PHYSICAL HARM OR PROPERTY DAMAGE (INCLUDING, FOR EXAMPLE, USE IN CONNECTION WITH ANY NUCLEAR, AVIONICS, LIFE SUPPORT OR OTHER LIFE CRITICAL APPLICATION). NVIDIA EXPRESSLY DISCLAIMS ANY EXPRESS OR IMPLIED WARRANTY OF FITNESS FOR SUCH HIGH RISK USES. NVIDIA SHALL NOT BE LIABLE TO CUSTOMER OR ANY THIRD PARTY, IN WHOLE OR IN PART, FOR ANY CLAIMS OR DAMAGES ARISING FROM SUCH HIGH RISK USES.

NVIDIA makes no representation or warranty that the product described in this document will be suitable for any specified use without further testing or modification. Testing of all parameters of each product is not necessarily performed by NVIDIA. It is customer’s sole responsibility to ensure the product is suitable and fit for the application planned by customer and to do the necessary testing for the application in order to avoid a default of the application or the product. Weaknesses in customer’s product designs may affect the quality and reliability of the NVIDIA product and may result in additional or different conditions and/or requirements beyond those contained in this document. NVIDIA does not accept any liability related to any default, damage, costs or problem which may be based on or attributable to: (i) the use of the NVIDIA product in any manner that is contrary to this document, or (ii) customer product designs.

Other than the right for customer to use the information in this document with the product, no other license, either expressed or implied, is hereby granted by NVIDIA under this document. Reproduction of information in this document is permissible only if reproduction is approved by NVIDIA in writing, is reproduced without alteration, and is accompanied by all associated conditions, limitations, and notices.

Trademarks

NVIDIA, the NVIDIA logo, TensorRT, NVIDIA Ampere, NVIDIA Hopper and NVIDIA Tesla are trademarks and/or registered trademarks of NVIDIA Corporation in the Unites States and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.

Copyright © 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.