Deploy container instances that use GPU resources

To run certain compute-intensive workloads on Azure Container Instances, deploy your container groups with GPU resources. The container instances in the group can access one or more NVIDIA Tesla GPUs while running container workloads such as CUDA and deep learning applications.

This article shows how to add GPU resources when you deploy a container group by using a YAML file or Resource Manager template. You can also specify GPU resources when you deploy a container instance using the Azure portal.


This feature is currently in preview, and some limitations apply. Previews are made available to you on the condition that you agree to the supplemental terms of use. Some aspects of this feature may change prior to general availability (GA).

Preview limitations

In preview, the following limitations apply when using GPU resources in container groups.

Region availability

Regions OS Available GPU SKUs
East US, West Europe, West US 2 Linux K80, P100, V100
Southeast Asia Linux P100, V100
Central India Linux V100
North Europe Linux K80

Support will be added for additional regions over time.

Supported OS types: Linux only

Additional limitations: GPU resources can't be used when deploying a container group into a virtual network.

About GPU resources

Count and SKU

To use GPUs in a container instance, specify a GPU resource with the following information:

  • Count - The number of GPUs: 1, 2, or 4.

  • SKU - The GPU SKU: K80, P100, or V100. Each SKU maps to the NVIDIA Tesla GPU in one the following Azure GPU-enabled VM families:

    SKU VM family
    K80 NC
    P100 NCv2
    V100 NCv3

Maximum resources per SKU

OS GPU SKU GPU count Max CPU Max Memory (GB) Storage (GB)
Linux K80 1 6 56 50
Linux K80 2 12 112 50
Linux K80 4 24 224 50
Linux P100 1 6 112 50
Linux P100 2 12 224 50
Linux P100 4 24 448 50
Linux V100 1 6 112 50
Linux V100 2 12 224 50
Linux V100 4 24 448 50

When deploying GPU resources, set CPU and memory resources appropriate for the workload, up to the maximum values shown in the preceding table. These values are currently larger than the CPU and memory resources available in container groups without GPU resources.


Default subscription limits (quotas) for GPU resources differ by SKU. The default CPU limits for the P100 and V100 SKUs are initially set to 0. To request an increase in an available region, please submit an Azure support request.

Things to know

  • Deployment time - Creation of a container group containing GPU resources takes up to 8-10 minutes. This is due to the additional time to provision and configure a GPU VM in Azure.

  • Pricing - Similar to container groups without GPU resources, Azure bills for resources consumed over the duration of a container group with GPU resources. The duration is calculated from the time to pull your first container's image until the container group terminates. It does not include the time to deploy the container group.

    See pricing details.

  • CUDA drivers - Container instances with GPU resources are pre-provisioned with NVIDIA CUDA drivers and container runtimes, so you can use container images developed for CUDA workloads.

    We support only CUDA 9.0 at this stage. For example, you can use the following base images for your Docker file:

YAML example

One way to add GPU resources is to deploy a container group by using a YAML file. Copy the following YAML into a new file named gpu-deploy-aci.yaml, then save the file. This YAML creates a container group named gpucontainergroup specifying a container instance with a K80 GPU. The instance runs a sample CUDA vector addition application. The resource requests are sufficient to run the workload.

additional_properties: {}
apiVersion: '2019-12-01'
name: gpucontainergroup
  - name: gpucontainer
          cpu: 1.0
          memoryInGB: 1.5
            count: 1
            sku: K80
  osType: Linux
  restartPolicy: OnFailure

Deploy the container group with the az container create command, specifying the YAML file name for the --file parameter. You need to supply the name of a resource group and a location for the container group such as eastus that supports GPU resources.

az container create --resource-group myResourceGroup --file gpu-deploy-aci.yaml --location eastus

The deployment takes several minutes to complete. Then, the container starts and runs a CUDA vector addition operation. Run the az container logs command to view the log output:

az container logs --resource-group myResourceGroup --name gpucontainergroup --container-name gpucontainer


[Vector addition of 50000 elements]
Copy input data from the host memory to the CUDA device
CUDA kernel launch with 196 blocks of 256 threads
Copy output data from the CUDA device to the host memory

Resource Manager template example

Another way to deploy a container group with GPU resources is by using a Resource Manager template. Start by creating a file named gpudeploy.json, then copy the following JSON into it. This example deploys a container instance with a V100 GPU that runs a TensorFlow training job against the MNIST dataset. The resource requests are sufficient to run the workload.

    "$schema": "",
    "contentVersion": "",
    "parameters": {
      "containerGroupName": {
        "type": "string",
        "defaultValue": "gpucontainergrouprm",
        "metadata": {
          "description": "Container Group name."
    "variables": {
      "containername": "gpucontainer",
      "containerimage": "microsoft/samples-tf-mnist-demo:gpu"
    "resources": [
        "name": "[parameters('containerGroupName')]",
        "type": "Microsoft.ContainerInstance/containerGroups",
        "apiVersion": "2019-12-01",
        "location": "[resourceGroup().location]",
        "properties": {
            "containers": [
              "name": "[variables('containername')]",
              "properties": {
                "image": "[variables('containerimage')]",
                "resources": {
                  "requests": {
                    "cpu": 4.0,
                    "memoryInGb": 12.0,
                    "gpu": {
                        "count": 1,
                        "sku": "V100"
        "osType": "Linux",
        "restartPolicy": "OnFailure"

Deploy the template with the az deployment group create command. You need to supply the name of a resource group that was created in a region such as eastus that supports GPU resources.

az deployment group create --resource-group myResourceGroup --template-file gpudeploy.json

The deployment takes several minutes to complete. Then, the container starts and runs the TensorFlow job. Run the az container logs command to view the log output:

az container logs --resource-group myResourceGroup --name gpucontainergrouprm --container-name gpucontainer


2018-10-25 18:31:10.155010: I tensorflow/core/platform/] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA
2018-10-25 18:31:10.305937: I tensorflow/core/common_runtime/gpu/] Found device 0 with properties:
name: Tesla K80 major: 3 minor: 7 memoryClockRate(GHz): 0.8235
pciBusID: ccb6:00:00.0
totalMemory: 11.92GiB freeMemory: 11.85GiB
2018-10-25 18:31:10.305981: I tensorflow/core/common_runtime/gpu/] Creating TensorFlow device (/device:GPU:0) -> (device: 0, name: Tesla K80, pci bus id: ccb6:00:00.0, compute capability: 3.7)
2018-10-25 18:31:14.941723: I tensorflow/stream_executor/] successfully opened CUDA library locally
Successfully downloaded train-images-idx3-ubyte.gz 9912422 bytes.
Extracting /tmp/tensorflow/input_data/train-images-idx3-ubyte.gz
Successfully downloaded train-labels-idx1-ubyte.gz 28881 bytes.
Extracting /tmp/tensorflow/input_data/train-labels-idx1-ubyte.gz
Successfully downloaded t10k-images-idx3-ubyte.gz 1648877 bytes.
Extracting /tmp/tensorflow/input_data/t10k-images-idx3-ubyte.gz
Successfully downloaded t10k-labels-idx1-ubyte.gz 4542 bytes.
Extracting /tmp/tensorflow/input_data/t10k-labels-idx1-ubyte.gz
Accuracy at step 0: 0.097
Accuracy at step 10: 0.6993
Accuracy at step 20: 0.8208
Accuracy at step 30: 0.8594
Accuracy at step 990: 0.969
Adding run metadata for 999

Clean up resources

Because using GPU resources may be expensive, ensure that your containers don't run unexpectedly for long periods. Monitor your containers in the Azure portal, or check the status of a container group with the az container show command. For example:

az container show --resource-group myResourceGroup --name gpucontainergroup --output table

When you're done working with the container instances you created, delete them with the following commands:

az container delete --resource-group myResourceGroup --name gpucontainergroup -y
az container delete --resource-group myResourceGroup --name gpucontainergrouprm -y

Next steps