Video Analytics Pipelines — DeepStream 9.1
DeepStream is NVIDIA's SDK for building accelerated intelligent video analytics (IVA) pipelines, and DeepStream 9.1 is the release that runs on Jetson Orin under JetPack 7.2. This page covers version pairing, install routes, decode limits, first-run expectations, headless RTSP output, and memory notes for the 8 GB Orin Nano Super Developer Kit.
1. Version pairing
DeepStream 9.1 ↔ JetPack 7.2 GA ↔ L4T 39.2 ↔ CUDA 13.2 ↔ TensorRT 10.16.1.7 ↔ GStreamer 1.24.2 (Docker image deepstream:9.1), as listed in the Platform and OS Compatibility table of the DeepStream Installation Guide.
DeepStream 8.0 and 9.0 listed AGX Thor only; 9.1 is the first 9.x release whose row includes Jetson Orin ("AGX Thor, Jetson Orin") — the row says "Jetson Orin" as a group; earlier rows (DS 6.3 to DS 7.1) named "Orin nano" explicitly. No 9.1 release note confirming Orin Nano specifically was found — treat support as implied by the group label (not yet confirmed). Baseline kit: JetPack 7.2.1 / L4T r39.2.1.
2. What this kit can decode
The Gst-nvvideo4linux2 decoder uses the NVDEC hardware engine and supports H.264, H.265, AV1, JPEG, and MJPEG. Published Orin Nano module capabilities:
| Capability | Specification |
|---|---|
| Video decode (H.265) | 1x 4K60 · 2x 4K30 · 5x 1080p60 · 11x 1080p30 |
| Video encode | No hardware encoder — "1080p30 supported by 1-2 CPU cores" |
| DLA · PVA | None |
Inference runs in the Gst-nvinfer plugin on TensorRT engines: FP16, FP32, and INT8 models (FP16 and INT8 are platform dependent); INT8 needs a calibration file. The plugin's enable-dla option has no engine to target on this module — the Orin Nano product page lists "DL Accelerator: -" and "Vision Accelerator: -".
At 8 GB: decoded frames, engines, and application memory share one pool, with no DLA to offload work to. The "30-stream" sample below decodes 30 1080p streams; this module's published decode capacity is 11x 1080p30 (H.265), so plan for fewer streams or lower resolution. There is also no hardware video encoder — encoded output (for example, RTSP streaming) runs on the CPU.
3. Install — Docker first
NVIDIA's guide says: "Recommended for new users: use Method 4 (Docker container) for the quickest, dependency-free setup." The four Jetson methods:
| Method | What it is |
|---|---|
| 1 — SDK Manager | Select DeepStreamSDK under "Additional SDKs" with the JetPack 7.2 GA components. |
| 2 — tar package | deepstream_sdk_v9.1.0_jetson.tbz2, a GitHub release asset. |
| 3 — Debian package | deepstream-9.1_9.1.0-1_arm64.deb. |
| 4 — Docker (recommended) | Jetson containers on NGC (nvcr.io). |
The Jetson containers are nvcr.io/nvidia/deepstream:9.1-samples-multiarch (reference applications, sample models and configs) and nvcr.io/nvidia/deepstream:9.1-triton-multiarch (plus devel libraries and Triton backends). Prerequisites: docker-ce, the NVIDIA Container Toolkit, an NGC account, and docker login nvcr.io (username $oauthtoken, password = your NGC API key).
Important: NVIDIA states: "The Jetson Docker containers are for deployment only. They do not support DeepStream software development within a container." Build applications natively on the kit and add your binaries to your own image.
In Docker, run user_additional_install.sh instead (see the EOS note below). The Triton container's "Failed to detect NVIDIA driver version" message is harmless.
Juxi tip: For a minimal host install, select only "Jetson OS" in SDK Manager, then run
sudo apt install docker.io,sudo apt install nvidia-container,sudo apt install nvidia-l4t-gstreamer, andsudo service docker restart.
4. Boost the clocks — with a kit-specific power mode
sudo nvpmodel -m 2
sudo jetson_clocksQuoted from the quickstart: "For Jetson Orin Nano modules, use sudo nvpmodel -m 2 instead of -m 0 to enable MAXN SUPER mode. For all other Jetson Orin modules (including Orin NX), use -m 0." Run these before running DeepStream applications. On a Super-configured 8 GB kit, the power modes are 15W (mode 0), 25W (mode 1, default), and MAXN_SUPER (mode 2); MAXN_SUPER exists only on Super-flashed units.
Attention: If 25W / MAXN SUPER is missing, or
nvpmodel -m 2reports a bad power mode, the unit was not flashed with the Super configuration. See Troubleshooting.
5. First run — TensorRT engines build on first use
cd /opt/nvidia/deepstream/deepstream-9.1/samples/configs/deepstream-app
deepstream-app -c source30_1080p_dec_infer-resnet_tiled_display.txtQuoted from the quickstart: for a model without an existing engine file, "it may take up to a few minutes (depending on the platform and the model) for the file generation and the application launch. For later runs, these generated engine files can be reused for faster loading." FPS metrics scroll in the terminal. The quickstart's "(~30 FPS for this configuration)" is the documentation's generic figure — it is not an Orin Nano measurement. If the application cannot create Gst elements, clear the cache and retry: rm ${HOME}/.cache/gstreamer-1.0/registry.aarch64.bin. Other sample configs cover USB and CSI cameras and tracking with secondary inference.
6. Headless operation with RTSP output
The quickstart documents how to run without a display: the default configs use the EGL-based nveglglessink renderer (type=2 in the [sink] groups), which requires a running X server. Add an RTSP output sink group instead — the [sink2] group in source30_1080p_dec_infer-resnet_tiled_display.txt is the example — and set enable=0 for the EGL sink group. The encoded RTSP output runs on the CPU (Section 2: no hardware encoder).
Juxi note: With RTSP streams, the application can get stuck reaching EOS (an
rtpjitterbufferissue). On bare metal, runupdate_rtpmanager.shin/opt/nvidia/deepstream/deepstream/once, after installing the Quickstart dependency packages. In Docker, runuser_additional_install.shinstead.
7. Memory planning for 8 GB
NVIDIA's memory-efficiency blog states: "Jetson Orin Nano 8 GB module, of the 8 GB physical DRAM, roughly 7.6 GB is usable after firmware and kernel reservations." CPU and GPU share this pool. Documented levers for DeepStream-style pipelines:
| Lever | Memory that may be reclaimed |
|---|---|
| Run bare metal instead of a container | Up to 70 MB |
| Switch from Python to C++ applications | Up to 84 MB |
| Disable Tiler/OSD and use FakeSink | Up to 258 MB |
| Total | 412 MB |
Disabling Tiler/OSD and using FakeSink "removes display stages needed for visualization but unnecessary in headless or production deployments. This saves memory, reduces GPU load, and improves throughput." This pairs with the headless RTSP path above; disabling the graphical desktop can free up to 865 MB. For the full 8 GB playbook, see Memory Efficiency on 8 GB.
What NVIDIA does not publish for this kit
The official DeepStream 9.1 performance page for Jetson covers two platforms only: Jetson AGX Thor and Jetson AGX Orin. No Orin Nano FPS figures are published; do not read AGX Orin rows as Orin Nano performance. For sizing, start from the decode capacity (Section 2), then lower stream count and resolution until the pipeline fits.
For the closest published datapoint, Ultralytics' Jetson benchmarks report YOLO26n on the Orin Nano Super at ~4.57 ms/image (~219 FPS) with a TensorRT FP16 engine and ~3.80 ms/image (~263 FPS) with INT8, at 640 input — vendor data, measured on JetPack 6.1-era software, not this kit's 7.2.1 stack; inference time excludes pre/post-processing. Per the same source, only the PyTorch, TorchScript, and TensorRT export formats use the GPU — other export formats run on the CPU.
Troubleshooting and further reading
- System-level issues (power modes, storage, display): Troubleshooting · Memory Efficiency on 8 GB.
- Models outside DeepStream: Local LLM Inference · official performance reference: DeepStream Performance.
Sources
- DeepStream Installation Guide (checked 2026-09-26)
- DeepStream Quickstart Guide (checked 2026-09-26)
- DeepStream Docker Containers (checked 2026-09-26)
- DeepStream Performance (checked 2026-09-26)
- Gst-nvvideo4linux2 (hardware decoder) (checked 2026-09-26)
- Gst-nvinfer (checked 2026-09-26)
- Jetson Orin modules — decode, encode, and accelerator specifications (checked 2026-09-26)
- Maximizing memory efficiency to run bigger models on NVIDIA Jetson (developer blog) (checked 2026-09-26)
- Jetson Linux r39.2 Developer Guide — Power and Performance (checked 2026-09-26)
- Ultralytics — NVIDIA Jetson guide (vendor benchmarks) (checked 2026-09-26)
Status: draft, pending review by cheny. Grounded in NVIDIA's official documentation as of the date listed; not yet verified on physical hardware by Juxi Technology.
NVIDIA® and Jetson™ are trademarks of NVIDIA Corporation. This page is published by Juxi Technology and is not an NVIDIA publication.