DEV Community

Cover image for First Look at NVIDIA Jetson Orin Nano Super - The Most Affordable Generative AI Supercomputer
Ajeet Singh Raina
Ajeet Singh Raina

Posted on • Edited on

First Look at NVIDIA Jetson Orin Nano Super - The Most Affordable Generative AI Supercomputer

#nvidia #docker #iot #developer

Table of Contents

NVIDIA has just reinvented edge computing with its latest offering - the Jetson Orin Nano Super Developer Kit. This isn't just an incremental update; it's a significant leap forward in bringing generative AI capabilities to the edge at an unprecedented price point of $249.

Comparing Jetson Nano Vs Jetson Orin Nano Super

The NVIDIA Jetson Orin Nano Super Developer Kit is a compact, yet powerful computer that redefines generative AI for small edge devices.

It delivers up to 67 TOPS of AI performance—a 1.7X improvement over its predecessor—to seamlessly run a wide variety of generative AI models, like vision transformers, large language models, vision-language models, and more.

It provides developers, students, and makers with the most affordable and accessible platform with the support of the NVIDIA AI software and a broad AI software ecosystem to democratize generative AI at the edge. Existing Jetson Orin Nano Developer Kit users can experience this performance boost with just a software upgrade, so everyone can
now unlock new possibilities with generative AI.

Learn about Jetson AI Lab

jetson_orin_1

Let's dive into the specs and see how it compares to its predecessors:

Feature Orin Nano Original Orin Nano Super Improvement
GPU Architecture NVIDIA Ampere (1024 CUDA cores, 32 Tensor cores) @ 635 MHz NVIDIA Ampere (1024 CUDA cores, 32 Tensor cores) @ 1020 MHz 1.6x GPU Clock
AI Performance 40 TOPS (Sparse) / 20 TOPS (Dense) 67 TOPS (Sparse) / 33 TOPS (Dense) 1.7x AI Performance
CPU 6-core Arm Cortex-A78AE @ 1.5 GHz 6-core Arm Cortex-A78AE @ 1.7 GHz 1.13x CPU Clock
Memory 8GB 128-bit LPDDR5 @ 68 GB/s 8GB 128-bit LPDDR5 @ 102 GB/s 1.5x Memory Bandwidth
Module Power 7W/15W 7W/15W/25W Additional Power Mode

How powerful is NVIDIA Jetson Orin Super?

jetson_orin_2

The most striking aspect of the Super variant is its performance improvements:

  • 1.7x increase in AI compute performance (67 TOPS vs 40 TOPS)
  • 1.5x increase in memory bandwidth (102 GB/s vs 68 GB/s)
  • Higher GPU and CPU clock speeds for better overall performance

Generative AI Capabilities

The NVIDIA Jetson™ platform runs the NVIDIA AI software stack, with a variety of available use-case-specific application frameworks. These include NVIDIA Isaac™ for robotics, NVIDIA Metropolis for vision AI, and NVIDIA Holoscan for sensor processing. You can save significant time with NVIDIA Omniverse™ Replicator for synthetic data generation (SDG) and NVIDIA TAO Toolkit for fine-tuning pretrained AI models from the NVIDIA® NGC™ catalog.

One of the most impressive aspects of the Orin Nano Super is its ability to run various types of generative AI models:
Large Language Models (LLMs):

Model Performance Gain
Llama-3.1 8B 1.37x
Llama 3.2 3B 1.55x
Qwen2.5 7B 1.53x
Gemma2 2B 1.63x
Gemma2 9B 1.28x
Phi 3.5 3B 1.54x
SmoLLM2 1.7B 1.57x

Vision Language Models (VLMs):

Model Performance Gain
VILA 1.5 3B 1.51x
VILA 1.5 8B 1.45x
LLAVA 1.6 7B 1.36x
Qwen2-VL-2B 1.57x
InternVL2.5-4B 2.04x
PaliGemma2-3B 1.58x
SmoLVLM-2B 1.59x

Vision Transformers

Model Performance Gain
clip-vit-base-patch32 1.60x
clip-vit-base-patch16 1.69x
DINOv2-base-patch14 1.68x
SAM2 base 1.43x
Grounding-DINO 1.52x
vit-base-patch16-224 1.61x
vit-base-patch32-224 1.60x

I/O and Connectivity

Interface Specification
Camera 2x MIPI CSI-2 22-pin Camera Connectors
PCIe M.2 Key M x4 PCIe Gen 3
Additional PCIe M.2 Key M x2 PCIe Gen3
Expansion M.2 Key E PCIe (x1), USB 2.0, UART, I2S, and I2C
USB 4x USB 3.2 Gen2 Type A + 1x Type C for Debug
Network 1x GbE Connector
Display DisplayPort 1.2 (+MST)
Storage microSD slot (UHS-1 cards up to SDR104 mode)
GPIO 40-Pin Expansion Header

Developer-Friendly Features

Feature Description
Software Stack Full support for TensorRT-LLM
Framework Compatibility Native compatibility with popular frameworks
Jetson Ecosystem Jetson Software Stack & Microservices support
Deployment Pre-built containers for rapid deployment

AI Development Tools

Tool Description
TensorRT Optimization Optimized inference using TensorRT
Quantization Support INT8/FP16 quantization support
Multi-Model Inference Ability to run multiple models simultaneously
Containerization Docker container support for easy deployment

Getting Started with Jetson Orin Super

Image8

This guide will walk you through setting up Ollama on your Jetson device, integrating it with Open WebUI, and configuring the system for optimal GPU utilization. Whether you're a developer or an AI enthusiast, this setup allows you to harness the full potential of LLMs right on your Jetson device.

Pre-requisite

What do you need?

Note: The NVIDIA Jetson Orin Nano Developer Kit doesn't support
HDMI, but it does have a DisplayPort output port. You can use an
adapter to connect the kit to a monitor or TV that only has HDMI.

  • NVMe SSD (instead of SD card)

Note: I highly recommend using NVMe SSD if you're serious about running
AI workload. Add NVMe SSD to the Jetson Board

image

  • WiFi Adapter
  • Wireless Keyboard
  • Wireless mouse
  • A Linux Laptop to install SDK Manager

Note: NVIDIA SDK Manager is required to flash Jetson Orin SDK card image to the board. Don't waste your time just writing it down to SD card directly. Save your time with SDK Manager.

Software

  • Download Jetson SD card image

Ensure that you download the latest JetPack 6.2 SDK from this link. Your Jetson Orin Nano Developer Kit comes with an old firmware flashed at the factory, which is NOT compatible with JetPack 6.x. Click here to download

  • Etcher installed on your local system

Download Jetson SDK using this link

Getting Started

  • Flashing the Jetson Board using SDK Manager - The Correct Way

To flash the Orin Nano using the SDK Manager, it must first be put > into “recovery mode.” To do that, attach a jumper or jumper wire between the FC_REC and GND pins (pins 2 and 3) on the underside of the Orin Nano card. Refer this blog to know more

Image8

jumper-pins

  • Ensure that SDK Manager detects the Jetson Board as well as NVMe SSD
  • Flash the NVMe SSD with the operating system

Image6

Step 1. Verify L4T Version

To check the L4T (Linux for Tegra) version on your NVIDIA Jetson device (e.g., Jetson Nano, Jetson Xavier), follow these steps:

Run the following command to retrieve your current L4T version.

head -n 1 /etc/nv_tegra_release

Results:

head -n 1 /etc/nv_tegra_release
# R36 (release), REVISION: 4.3, GCID: 38968081, BOARD: generic, EABI: aarch64, DATE: Wed Jan  8 01:49:37 UTC 2025
Enter fullscreen mode Exit fullscreen mode

Here are the list of supported L4T versions:

  • 35.3.1
  • 35.4.1
  • 35.5.0
  • 36.3.0

If your L4T version does not match the supported versions listed above, you may need to re-flash the system on your NVIDIA Jetson device. You might need to use SDK Manager on another computer to re-flash the device. You can download the SDK Manager and follow the tutorial from the official NVIDIA Developer site.

Step 2. Keep apt up to date: 

   sudo apt update && sudo apt upgrade
Enter fullscreen mode Exit fullscreen mode

Step 3. Install jetpack: 

   sudo apt install jetpack
Enter fullscreen mode Exit fullscreen mode

Step 4. Add users

Add your user to the docker group and restart the Docker service to apply the change:

   sudo usermod -aG docker $USER && \
   newgrp docker && \
   sudo systemctl daemon-reload && \
   sudo systemctl restart docker
Enter fullscreen mode Exit fullscreen mode

Step 5. Install jetson-examples: 

   pip3 install jetson-examples
Enter fullscreen mode Exit fullscreen mode

Step 6. Reboot system 

   sudo reboot
Enter fullscreen mode Exit fullscreen mode

Step 7. Install Ollama 

   reComputer run ollama
Enter fullscreen mode Exit fullscreen mode

Optional: If you run the above command via ssh and encounter the error command not found: reComputer, you can resolve this by executing the following command:

   source ~/.profile
Enter fullscreen mode Exit fullscreen mode

Step 8. Run a model

I was able to run DeepSeek R1 model and the overall result was quite impressive.

asciicast

Step 9. Install models (e.g. llama3.2) from Ollama Library 

ollama pull llama3.2
Enter fullscreen mode Exit fullscreen mode

Step 9. Install and run Open WebUI through Docker

Jetson Orin comes with Docker installed by default. As of writing, I had the following Docker binaries installed by default:

ajeetraina@ajeetraina-desktop:~$ sudo docker version
[sudo] password for ajeetraina:
Client: Docker Engine - Community
 Version:           28.0.1
 API version:       1.48
 Go version:        go1.23.6
 Git commit:        068a01e
 Built:             Wed Feb 26 10:41:16 2025
 OS/Arch:           linux/arm64
 Context:           default

Server: Docker Engine - Community
 Engine:
  Version:          28.0.1
  API version:      1.48 (minimum version 1.24)
  Go version:       go1.23.6
  Git commit:       bbd0a17
  Built:            Wed Feb 26 10:41:16 2025
  OS/Arch:          linux/arm64
  Experimental:     false
 containerd:
  Version:          1.7.25
  GitCommit:        bcc810d6b9066471b0b6fa75f557a15a1cbf31bb
 runc:
  Version:          1.2.4
  GitCommit:        v1.2.4-0-g6c52b3f
 docker-init:
  Version:          0.19.0
  GitCommit:        de40ad0
Enter fullscreen mode Exit fullscreen mode 
docker run -d -p 3000:8080 --gpus all --add-host=host.docker.internal:host-gateway -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:cuda
Enter fullscreen mode Exit fullscreen mode

Step 10. Install and run Open WebUI through docker

Once the installation is finished, you can access the GUI by visiting YOUR_SERVER_IP:3000 in your browser.

Access the API endpoints by navigating to YOUR_SERVER_IP/ollama/docs#/. For comprehensive documentation, please refer to the official resources: the Ollama API Documentation (recommended) and Open WebUI API Endpoints.

Using GPU

This installation method uses a single container image that bundles Open WebUI with Ollama, allowing for a streamlined setup via a single command. Choose the appropriate command based on your hardware setup:

sudo docker run -d -p 3000:8080 --gpus=all -v ollama:/root/.ollama -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:ollama
Enter fullscreen mode Exit fullscreen mode

Using CPU only

For CPU Only: If you're not using a GPU, use this command instead:

sudo docker run -d -p 3000:8080 -v ollama:/root/.ollama -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:ollama
Enter fullscreen mode Exit fullscreen mode

Both commands facilitate a built-in, hassle-free installation of both Open WebUI and Ollama, ensuring that you can get everything up and running swiftly.

Once configured, Open WebUI can be accessed at http://localhost:3000, while Ollama operates at http://localhost:11434. This setup provides a seamless and GPU-accelerated environment for running and managing LLMs locally on NVIDIA Jetson devices.

Setting up TensorRT LLM

TensorRT-LLM is essentially a specialized tool that makes large language models (like ChatGPT) run much faster on NVIDIA hardware.

Think of it this way: If a regular language model is like a car engine that can get you from point A to point B, TensorRT-LLM is like a high-performance tuning kit that makes that same engine much more efficient and powerful.
In simple terms:

  • It takes AI language models that would normally be slow to respond and optimizes them to run much faster.
  • It's specifically designed for NVIDIA hardware like your Jetson Orin, helping it punch above its weight when running complex AI models.
  • It works by restructuring the model to eliminate inefficiencies, similar to how a mechanic might tune a car engine to get better performance without changing the basic design.
  • The end result is that AI models that might be too slow to be practical on your device can now run with reasonable response times.

TensorRT-LLM is a high-performance LLM inference library with advanced quantization, attention kernels, and paged KV caching. Initial support for building TensorRT-LLM from source for JetPack 6.1 has been included in the v0.12.0-jetson branch of the TensorRT-LLM repo for Jetson AGX Orin.

I recommend NVMe SSD highly recommended for storage speed and space

  • 18.5GB for tensorrt_llm container image
  • Space for models ( >10GB ) 
git clone https://github.com/dusty-nv/jetson-containers
bash jetson-containers/install.sh
Enter fullscreen mode Exit fullscreen mode

Building TensorRT-LLM Engine for Llama

You can find the steps for converting Llama to TensorRT-LLM under examples/llama in the repo, and also in the documentation . This script will automate the process for Llama-7B with INT4 quantization applied, and run some generation and performance checks on the model:

First, get the Hugging Face API token , pass it to the command and run it:

First, create the directory with the proper permissions:

sudo mkdir -p /tmp/llama2-config
sudo chmod 777 /tmp/llama2-config
Enter fullscreen mode Exit fullscreen mode

Then try creating the files again:

cat > /tmp/llama2-config/config.json << 'EOF'
{
  "_name_or_path": "meta-llama/Llama-2-7b-chat-hf",
  "architectures": ["LlamaForCausalLM"],
  "attn_implementation": "sdpa",
  "bos_token_id": 1,
  "eos_token_id": 2,
  "hidden_act": "silu",
  "hidden_size": 4096,
  "initializer_range": 0.02,
  "intermediate_size": 11008,
  "max_position_embeddings": 4096,
  "model_type": "llama",
  "num_attention_heads": 32,
  "num_hidden_layers": 32,
  "num_key_value_heads": 32,
  "pretraining_tp": 1,
  "rms_norm_eps": 1e-05,
  "rope_scaling": null,
  "tie_word_embeddings": false,
  "torch_dtype": "float16",
  "transformers_version": "4.31.0",
  "use_cache": true,
  "vocab_size": 32000
}
EOF
Enter fullscreen mode Exit fullscreen mode

Also create the tokenizer config:

cat > /tmp/llama2-config/tokenizer_config.json << 'EOF'
{
  "add_bos_token": true,
  "add_eos_token": false,
  "bos_token": {
    "__type": "AddedToken",
    "content": "<s>",
    "lstrip": false,
    "normalized": true,
    "rstrip": false,
    "single_word": false
  },
  "clean_up_tokenization_spaces": false,
  "eos_token": {
    "__type": "AddedToken",
    "content": "</s>",
    "lstrip": false,
    "normalized": true,
    "rstrip": false,
    "single_word": false
  },
  "model_max_length": 1000000000000000019884624838656,
  "tokenizer_class": "LlamaTokenizer",
  "unk_token": {
    "__type": "AddedToken",
    "content": "<unk>",
    "lstrip": false,
    "normalized": true,
    "rstrip": false,
    "single_word": false
  }
}
EOF
Enter fullscreen mode Exit fullscreen mode

It’s time to convert the model:

jetson-containers run \
  -e HUGGINGFACE_TOKEN=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 \
  -e FORCE_BUILD=on \
  -v /tmp/llama2-config:/tmp/llama2-config \
  dustynv/tensorrt_llm:0.12-r36.4.0 \
  bash -c "cd /opt/TensorRT-LLM/examples/llama && python3 convert_checkpoint.py \
  --model_dir /tmp/llama2-config \
  --output_dir /data/models/tensorrt_llm/Llama-2-7b-chat-hf-gptq \
  --dtype float16 \
  --quant_ckpt_path /data/models/huggingface/models--TheBloke--Llama-2-7B-Chat-GPTQ/snapshots/d5ad9310836dd91b6ac6133e2e47f47394386cea/model.safetensors \
  --use_weight_only \
  --weight_only_precision int4_gptq \
  --group_size 128 \
  --per_group"
Enter fullscreen mode Exit fullscreen mode

Great! The conversion is running successfully now. You can see it's processed 550 iterations at a rate of 4.83 iterations per second, which is good progress.

The command is working correctly, and the model is being converted to the TensorRT-LLM format. This process will continue for a while as it works through all the model layers and weights to optimize them for your Jetson hardware.

To run the conversion process without entering the container, you'll need to create a single command that handles the complete workflow. I'll create a script that you can save and run:

#!/bin/bash

# Save this as convert_llama_gptq.sh and make it executable with: chmod +x convert_llama_gptq.sh

# Run the container with all the necessary steps in sequence
jetson-containers run \
  -it \
  dustynv/tensorrt_llm:0.12-r36.4.0 \
  bash -c '
    # Step 1: Install GPTQ model support
    pip install gptqmodel

    # Step 2: Download the complete model with configs
    cd /tmp
    python3 -c "from transformers import AutoModelForCausalLM; AutoModelForCausalLM.from_pretrained('\''TheBloke/Llama-2-7B-Chat-GPTQ'\'', trust_remote_code=True)"

    # Step 3: Find the snapshot directory
    SNAPSHOT_DIR=$(find /root/.cache/huggingface/hub/models--TheBloke--Llama-2-7B-Chat-GPTQ/snapshots -type d | head -1)
    echo "Using model directory: $SNAPSHOT_DIR"

    # Step 4: Run the conversion
    cd /opt/TensorRT-LLM/examples/llama
    python3 convert_checkpoint.py \
      --model_dir $SNAPSHOT_DIR \
      --output_dir /data/models/tensorrt_llm/Llama-2-7b-chat-hf-gptq \
      --dtype float16 \
      --use_weight_only \
      --weight_only_precision int4_gptq \
      --group_size 128 \
      --per_group
Enter fullscreen mode Exit fullscreen mode

Save this script to a file, make it executable, and run it. The entire process will be executed within the container in sequence without you needing to enter the container manually.
The script:

  • Installs the GPTQ model support
  • Downloads the complete model with all configs
  • Finds the actual snapshot directory where the model was downloaded
  • Runs the conversion with the proper directory

Note that this will still take a significant amount of time to complete, especially the conversion step. You'll see the output in your terminal as the process runs.

#!/bin/bash

# Save this as convert_llama_gptq.sh and make it executable with: chmod +x convert_llama_gptq.sh

# Run the container with all the necessary steps in sequence
jetson-containers run \
  -it \
  dustynv/tensorrt_llm:0.12-r36.4.0 \
  bash -c '
    # Step 1: Install required packages
    pip install auto-gptq

    # Step 2: Create a config directory with required files
    mkdir -p /tmp/llama_config

    # Create config.json
    cat > /tmp/llama_config/config.json << EOF
{
  "_name_or_path": "meta-llama/Llama-2-7b-chat-hf",
  "architectures": ["LlamaForCausalLM"],
  "bos_token_id": 1,
  "eos_token_id": 2,
  "hidden_act": "silu",
  "hidden_size": 4096,
  "initializer_range": 0.02,
  "intermediate_size": 11008,
  "max_position_embeddings": 4096,
  "model_type": "llama",
  "num_attention_heads": 32,
  "num_hidden_layers": 32,
  "num_key_value_heads": 32,
  "rms_norm_eps": 1e-05,
  "rope_scaling": null,
  "tie_word_embeddings": false,
  "transformers_version": "4.31.0",
  "use_cache": true,
  "vocab_size": 32000
}
EOF

    # Step 3: Run the conversion
    cd /opt/TensorRT-LLM/examples/llama
    python3 convert_checkpoint.py \
      --model_dir /tmp/llama_config \
      --output_dir /data/models/tensorrt_llm/Llama-2-7b-chat-hf-gptq \
      --dtype float16 \
      --quant_ckpt_path /data/models/huggingface/models--TheBloke--Llama-2-7B-Chat-GPTQ/snapshots/d5ad9310836dd91b6ac6133e2e47f47394386cea/model.safetensors \
      --use_weight_only \
      --weight_only_precision int4_gptq \
      --group_size 128 \
      --per_group
Enter fullscreen mode Exit fullscreen mode

Conclusion

The Jetson Orin Nano Super Developer Kit represents a significant milestone in edge AI computing. It brings datacenter-class AI capabilities to the edge at an unprecedented price point, making it an ideal platform for developers, researchers, and businesses looking to deploy advanced AI applications at the edge.

The combination of increased AI performance, enhanced memory bandwidth, and broad model support makes it a compelling choice for anyone serious about edge AI development. At $249, it's not just a product - it's a revolution in accessible AI computing.

In the next blog post, we will deep dive further into model optimisation. Stay tuned.

References

Top comments (10)

Collapse
 
ajeetraina profile image
Ajeet Singh Raina • • Edited

Quick Update[17 feb 2025]::

I highly recommend using SSD instead of microSD card for AI workloads. I ordered "Crucial P3 500GB PCIe 3.0 3D NAND NVMe M.2 SSD, up to 3500MB/s - CT500P3SSD8" from Amazon. Will keep you all posted on the performance.

Expecting NVMe SSDs to reach atleast 3500 MB/s+ (35x faster than SD cards). Need a faster boot times, quicker model loading, and improved system responsiveness.

Collapse
 
ajeetraina profile image
Ajeet Singh Raina • • Edited

Quick Update: 22- Feb 2025:

Image description

My containers just got a power boost! ⚡ Upgraded my Jetson Orin Nano with a Crucial P3 NVMe SSD—more speed, more storage, and extra energy for AI workloads! 🚀 #JetsonOrin #EdgeAI #Containers #NVMePower

Collapse
 
nathaniel_palmieri_9d5830 profile image
Nathaniel Palmieri •

Thank you for your insights.

Collapse
 
exaptedai777 profile image
Exapted Aint •

Will this run any GPU-requiring code or is it limited LLM installs only?

Collapse
 
ajeetraina profile image
Ajeet Singh Raina • • Edited

It’s not just restricted to LLM but run any GPU requiring code.

Collapse
 
nkko profile image
Nikola Balic •

Great, would be even better to have some tests and results like tokens/sec for some models.

Collapse
 
click2install profile image
click2install •

Triton makes it super easy to spin up a model using various backends.

Collapse
 
ajeetraina profile image
Ajeet Singh Raina •

That's on my wishlist:

docker run --gpus all --rm -p 8000:8000 -p 8001:8001 -p 8002:8002 \
    -v /path/to/model/repository:/models \
    nvcr.io/nvidia/tritonserver:23.12-py3 \
    tritonserver --model-repository=/models
Enter fullscreen mode Exit fullscreen mode
Collapse
 
darkpowerxo profile image
Sam Abtahi • • Edited

Anyone knows where to buy this in canada? Other than the the 999$ option on amazon instead of the 250 ?

Collapse
 
ajeetraina profile image
Ajeet Singh Raina • • Edited

Did you check https://www.amazon.ca/dp/B0BZJTQ5YP?ref=cm_sw_r_cso_cp_apin_dp_RBZDTNN21DGZNMPPJ3GW&ref_=cm_sw_r_cso_cp_apin_dp_RBZDTNN21DGZNMPPJ3GW&social_share=cm_sw_r_cso_cp_apin_dp_RBZDTNN21DGZNMPPJ3GW&starsLeft=1 Look like it’s out of stock.

Read next

dipe_ profile image

Deploying a FastAPI Application with CI/CD Pipeline: HNG Task 2

Precious Ogundipe -

bobbyiliev profile image

How to Build Windows Docker Images with GitHub Actions

Bobby Iliev -

code42cate profile image

This Week in Docker: AI, AI, AI!

Jonas Scholz -

thatanjan profile image

How to Install Docker on Arch Linux

Anjan Shomodder -