Nvidia Jetson installation

Download SD card image

https://developer.download.nvidia.com/assets/embedded/downloads/jetson-nano-4gb-jp441-sd-card-image/jetson-nano-4gb-jp441-sd-card-image.zip

Create SD card for NVIDIA Jetson Nano

This command extracts the downloaded zip image and writes it directly to the SD card device using dd. Make sure to replace /dev/mmcblk0 with the correct device path for your SD card.

unzip -p  ~/Downloads/jetson-nano-4gb-jp441-sd-card-image.zip | sudo /bin/dd of=/dev/mmcblk0  bs=1M status=progress

Update Docker runtime

To enable GPU access for Docker containers on the Jetson, you need to set the NVIDIA runtime as the default Docker runtime. Back up the existing configuration, then update daemon.json and restart Docker.

sudo cp /etc/docker/daemon.json  /etc/docker/daemon.json.orig
ubuntu@k3s-jetson-1:~$
ubuntu@k3s-jetson-1:~$
ubuntu@k3s-jetson-1:~$ sudo vim /etc/docker/daemon.json
ubuntu@k3s-jetson-1:~$ cat  /etc/docker/daemon.json
{
    "runtimes": {
        "nvidia": {
            "path": "nvidia-container-runtime",
            "runtimeArgs": []
        }
    },
    "default-runtime": "nvidia"
}

sudo systemctl restart docker

ubuntu@k3s-jetson-1:~$ sudo docker info | grep Def
Default Runtime: nvidia

Try to run a Docker container with TensorFlow

Verify that the GPU is accessible by running the official NVIDIA L4T TensorFlow container. This command starts an interactive Python session where you can confirm TensorFlow detects the GPU.

sudo docker run -it --rm --runtime nvidia --network host nvcr.io/nvidia/l4t-tensorflow:r32.4.3-tf2.2-py3 python3

Python 3.6.9 (default, Apr 18 2020, 01:56:04)
[GCC 8.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>>
>>>
>>> 1
1
>>> 2
2
>>> import tensorflow as tf
2020-11-19 12:04:09.268672: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.2
>>> print(tf.__version__)
2.2.0


>>> tf.config.list_physical_devices('GPU')
2020-11-19 12:05:17.191282: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcuda.so.1
2020-11-19 12:05:17.206274: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:948] ARM64 does not support NUMA - returning NUMA node zero
2020-11-19 12:05:17.206430: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1561] Found device 0 with properties:
pciBusID: 0000:00:00.0 name: NVIDIA Tegra X1 computeCapability: 5.3
coreClock: 0.9216GHz coreCount: 1 deviceMemorySize: 3.87GiB deviceMemoryBandwidth: 194.55MiB/s
2020-11-19 12:05:17.206513: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.2
2020-11-19 12:05:17.267906: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcublas.so.10
2020-11-19 12:05:17.348069: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcufft.so.10
2020-11-19 12:05:17.471141: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcurand.so.10
2020-11-19 12:05:17.655922: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcusolver.so.10
2020-11-19 12:05:17.739489: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcusparse.so.10
2020-11-19 12:05:17.741373: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudnn.so.8
2020-11-19 12:05:17.742026: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:948] ARM64 does not support NUMA - returning NUMA node zero
2020-11-19 12:05:17.742338: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:948] ARM64 does not support NUMA - returning NUMA node zero
2020-11-19 12:05:17.742455: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1703] Adding visible gpu devices: 0
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

You can also deploy a TensorFlow pod on a K3s cluster running on the Jetson. The following YAML defines a simple pod that keeps running so you can exec into it for interactive GPU testing.

cat pod.yaml
apiVersion: v1
kind: Pod
metadata:
  name: tensorflow
spec:
  containers:
  - name: tf
    image: nvcr.io/nvidia/l4t-tensorflow:r32.4.3-tf2.2-py3
    command: [ "/bin/bash", "-c", "--" ]
    args: [ "while true; do sleep 30; done;" ]

ubuntu@k3s-jetson-1:~$ logout
kubectl apply -f pod.yaml
pod/tensorflow created

kubectl get pods
NAME         READY   STATUS    RESTARTS   AGE
tensorflow   1/1     Running   0          5s

kubectl get pods -o wide
NAME         READY   STATUS    RESTARTS   AGE   IP          NODE           NOMINATED NODE   READINESS GATES
tensorflow   1/1     Running   0          9s    10.42.1.4   k3s-jetson-1   <none>           <none>