Deepstream 9.1 Production Base Image

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1
  • Hardware Platform: RTX5070Ti/Jetson Orin AGX 64GB
  • DeepStream Version 9.1
  • JetPack Version (valid for Jetson only) 7.2.1
  • TensorRT Version 10.16.x
  • NVIDIA GPU Driver Version (valid for GPU only) 610 open
  • Question

What is the recommended minimal size docker base image to deploy a compiled C++ binary as part of a Docker Image.
There seem to be only two images supported:

deepstream-9.1:triton-multiarch:
This seems to be the “large” one with still > 12 Gigabytes, containing all samples and dev dependencies

deepstream-9.1:samples:
this is a little bit smaller but still huge.

I already created a split approach having a base image with additional installs like gstreamer-rs plugins for webrtc, but this base image is 18 Gigabytes, then my release image only adds the compiled binary ~5MB to it.
If i ever need to update the base image accross the fleet downloading ~20 Gigabytes per machine is a no-go, as in deepstream 9.2 updates etc.. This size seems like it is bloated and contains thinges i will never use in production.

What ways does nvidia recommend to deploy deepstream based applications in 2026?

3

You could consider building your own Docker image from dockerfile.

If you want to minimize the image size, you could remove unused video/.so files from the image.

Noting that you’re using Orin. I don’t recommend Docker deployment because Docker images depend on the Jetpack version; host deployment is generally a better recommendation.

4

For DeepStream 9.1, please refer to DeepStream/deepstream_dockers at main · NVIDIA/DeepStream for the docker files.

5

before i go through the effort to create my own deepstream docker file.
can you estimate the reduction in size.
Only saving on some .so files seems like its not worth it.
I was talking like tools that are never used, nsight systems for example.
Where does the 14 Gigabyte difference between the base cuda image and the deepstream images come from?

6

if docker deployment is not recommendet, why are you providing Dockerfiles specifically for jetson?
I dont see a different way, as our services are deployed using k8s

7

No relevant data is currently available.
The increased size includes the gstreamer base library, TensorRT, tritonserver(for nvinferserver), deepstream libraries, deepstream sdk test videos, etc.

This is just a suggestion that it’s not recommended. If you’re using Kubernetes, please continue to use container deployment.

8

my pipeline is not using tritonserver.
also the test videos are not needed as well, are these also contained in the samples docker image?
if so, then i will give it a shot to build my own base image

9

Compared to the TensorRT base image, the samples-multiarch image adds only about 2 GB. Libraries related to Triton Inference Server have been removed from samples-multiarch;
Disk USAGE refers to the space occupied after extraction; this is related to Docker image layers, This is not something DeepStream can handle.

IMAGE                                                     ID             DISK USAGE   CONTENT SIZE 
nvcr.io/nvidia/cuda:13.0.0-tensorrt-runtime-ubuntu24.04   198208c45877       9.69GB         4.19GB 
nvcr.io/nvidia/deepstream:9.1-samples-multiarch           4f80b374e4a5       16.2GB         6.39GB

The /opt/nvidia/deepstream/deepstream-9.1/samples directory within the image—containing models and test video files—accounts for approximately 624 MB.
Whether the other directories can be deleted depends on your application.

root@d6ec4519c615:/opt/nvidia/deepstream/deepstream-9.1/samples# ls
configs  models  prepare_classification_test_video.sh  prepare_ds_triton_model_repo.sh  prepare_ds_triton_tao_model_repo.sh  streams  triton_backend_setup.sh  triton_model_repo  triton_tao_model_repo  trtis_model_repo
root@d6ec4519c615:/opt/nvidia/deepstream/deepstream-9.1/samples# du -sh .
624M	.