Creating and deleting TPUs

Running a Machine Learning (ML) model requires a Compute Engine VM and Cloud TPU resources. This page describes how to manage these resources using:

• The ctpu utility, which provides a CLI specifically designed for managing Cloud TPU resources
• The gcloud command-line tool, which provides the primary CLI to Google Cloud Platform (GCP)
• The Cloud Console, which provides an integrated management console for your GCP resources.

Prerequisites

To run these procedures, you need to have a Google Cloud Platform (GCP) project set up. If you don't have a project, see Creating and managing projects to set one up.

Setting up a Compute Engine VM

Note: Use the same name and zone for the VM and TPU resources.

ctpu

The ctpu utility can create the Compute Engine VM and Cloud TPU resources together or separately. In this procedure, ctpu is used to only create the Compute Engine VM.

1. In the Cloud shell, run the following command to create a Compute Engine VM:

$ ctpu up --vm-only --zone=zone [optional: --name --machine-type --disk-size-gb]

Parameter	Description
vm-only	Create only the Compute Engine VM.
zone	The zone where you plan to create your Cloud TPU. For example, us-central1-b.
name	Specify a name for the Compute Engine VM. If a name is not specified, the default is your username.
machine-type	This is the type of machine to use for the VM. See machine types for the supported machine types. The default is n1-standard-1.
disk-size-gb	The disk size in GB for the VM. For example, 300GB. If a size is not specified, it defaults to 20GB.

See the ctpu Reference for all of the ctpu options.

gcloud commands

Use gcloud commands to interact with Google Cloud in the Cloud shell.

1. If you are not using the Cloud Shell as your command interface, set up the gcloud command-line tool by installing the Cloud SDK for your operating system.
2. Configure gcloud to use your project.

gcloud config set project project-name

3. Specify the zone where you plan to create your Compute Engine VM. For example, us-central1-b.

$ gcloud config set compute/zone zone

4. Launch a Compute Engine virtual machine.

   Since you specified the zone in the previous command, the VM instance is created in that zone.

   gcloud compute instances create vm-name \
   --machine-type=machine-type \
   --image-project=ml-images \
   --image-family=tensorflow-vm-image \
   --boot-disk-size=boot-disk-size \
   --scopes=cloud-platform
   

Parameter	Description
vm-name	Specify a name for the Compute Engine VM.
machine-type	This is the type of machine to use for the VM. See machine types for the supported machine types.
image-project	The project against which all image and image family references will be resolved. Use ml-images.
image-family	This is the TensorFlow family of the image that the boot disk will be initialized with, for example, tf-1-15 or tf-2-1.
boot-disk-size	The boot disk size in GB for the VM. For example, 300GB. If a size is not specified, it defaults to 20GB.
scopes	Use cloud-platform.

This will generate output similar to the following:

NAME         ZONE           MACHINE_TYPE    PREEMPTIBLE INTERNAL_IP  EXTERNAL_IP    STATUS
demo-vm-tpu  us-central1-b  n1-standard-1               10.138.0.2   35.247.15.162  RUNNING

5. Remotely connect to your Compute Engine VM:

   $ gcloud compute ssh vm-name --zone=zone
   

Console

From the Google Cloud Console, create your VM and establish remote access to it.

1. Select Compute Engine > VM instances from the left-hand navigation bar and click CREATE INSTANCE.
2. From the top menu bar on the Create an instance page, select Create an instance and specify an instance name, the region, and a machine type.

Parameter	Description
name	Specifies the name of the Compute Engine VM. You can specify any instance name, but use the same one for both the VM instance and the Cloud TPU.
region	If you have set up a Cloud Storage bucket, this should match the Location of the bucket.
machine type	Specifies the machine type to use for your Compute Engine VM. Select a machine type from the drop down menu.

3. Click the Boot Disk > Change button. This will bring up a new window where you can choose your VM operating system and version.
4. In the Operating system pull down, select Deep Learning on Linux.
5. Under Version select the version of Debian GNU/Lunix with the TensorFlow version you want to run, for example, TF2-1.
6. At the bottom of the Boot Disk page, click Select to complete the Operating System and Version selection.
7. Under Identity and API access > Access scopes, select Allow full access to all Cloud APIs.
8. Go to Compute Engine > VM instances. Find the instance with your VM name, and click SSH to connect to it.

Setting up a Cloud TPU

Note: Use the same name and zone for the VM and TPU resources.

Set up your Compute Engine VM using the VM setup procedure before setting up your Cloud TPU. You can allocate and start your TPU resources using the ctpu utility, gcloud commands, or the Cloud Console.

ctpu

Run the following command in the Cloud shell to create your Cloud TPU.

$ ctpu up --tpu-only --name=tpu-name --zone=zone [optional: --tpu-size]

Parameter	Description
name	Specifies the name of the Cloud TPU. Use the same name as you used for the Compute Engine VM.
zone	The zone where you plan to create your Cloud TPU. This should be the same zone you used for the Compute Engine VM. For example, us-central1-b.
tpu-size	This is the TPU type to use. The default is v2-8. See Types and zones for the supported TPU types and zones.

gcloud commands

The Cloud SDK is a set of tools that you can use to interact with Google Cloud in the Cloud shell.

1. Install the gcloud command-line tool via the Cloud SDK.
2. Use the gcloud command-line tool to specify your Google Cloud project:

   $ gcloud config set project project-name
   

3. Specify the zone where you plan to create your Cloud TPU resource. This should be the same zone you used for the Compute Engine VM. For example, us-central1-b.

   $ gcloud config set compute/zone zone
   

   Cloud TPU is available in the following zones:

   US

   TPU type (v2)	TPU v2 cores	Total TPU memory	Region/Zone
   v2-8	8	64 GiB	us-central1-b us-central1-c us-central1-f
   v2-32	32	256 GiB	us-central1-a
   v2-128	128	1 TiB	us-central1-a
   v2-256	256	2 TiB	us-central1-a
   v2-512	512	4 TiB	us-central1-a
   TPU type (v3)	TPU v3 cores	Total TPU memory	Available zones
   v3-8	8	128 GiB	us-central1-a us-central1-b us-central1-f

   Europe

   TPU type (v2)	TPU v2 cores	Total TPU memory	Region/Zone
   v2-8	8	64 GiB	europe-west4-a
   v2-32	32	256 GiB	europe-west4-a
   v2-128	128	1 TiB	europe-west4-a
   v2-256	256	2 TiB	europe-west4-a
   v2-512	512	4 TiB	europe-west4-a
   TPU type (v3)	TPU v3 cores	Total TPU memory	Available zones
   v3-8	8	128 GiB	europe-west4-a
   v3-32	32	512 GiB	europe-west4-a
   v3-64	64	1 TiB	europe-west4-a
   v3-128	128	2 TiB	europe-west4-a
   v3-256	256	4 TiB	europe-west4-a
   v3-512	512	8 TiB	europe-west4-a
   v3-1024	1024	16 TiB	europe-west4-a
   v3-2048	2048	32 TiB	europe-west4-a

   Asia Pacific

   TPU type (v2)	TPU v2 cores	Total TPU memory	Region/Zone
   v2-8	8	64 GiB	asia-east1-c

4. Create a new Cloud TPU resource.
   Note: Billing begins as soon as the Cloud TPU is created, and continues until the time it is deleted. Check the Cloud TPU pricing page to estimate your costs.

   Since you specified the zone in the previous command, the Cloud TPU is created in that zone.

   $ gcloud compute tpus create tpu-name \
         --network=network-ID or default \
         --accelerator-type=tpu-type \
         --version=2.1
   

   Parameter	Description
   tpu-name	Specifies the name of the Cloud TPU. Use the same name you used for the Compute Engine VM name.
   network	If you know your network ID, use that, otherwise enter default.
   accelerator-type	This is your TPU type. See TPU types for the supported TPU types for your zone.
   version	The TensorFlow version to use with your Cloud TPU. This must be the same version of TensorFlow used to write your model.

   This will generate output similar to the following:

   NAME         ZONE           ACCELERATOR_TYPE NETWORK_ENDPOINT  NETWORK  RANGE         STATUS
   demo-vm-tpu  us-central1-b  v2-8             10.240.1.2:8470   default  10.240.1.0/29 READY
   

5. Remotely connect to your Compute Engine VM:

   $ gcloud compute ssh vm-name
   

6. Create an environment variable containing the name of your TPU:

   $ export TPU_NAME=tpu-name
   

Console

Create, start and connect to your Cloud TPU.

Note: Cloud TPU charges begin when you create your Cloud TPU.
1. Go to Compute Engine > TPUs on the left-hand navigation bar and click CREATE TPU NODE.
2. On the Create a Cloud TPU page use the menu pulldowns to specify the TPU name, the zone, TPU type, TPU software version, and network.

Parameter	Description
name	Specifies the name of the Cloud TPU. Use the same name you used for the Compute Engine VM name.
zone	The zone where you plan to create your Cloud TPU. For example, us-central1-b.
TPU type	This is your TPU type. See TPU types for the supported TPU types for your zone.
TPU software version	The TensorFlow or PyTorch version to use with your Cloud TPU. This must be the same version of TensorFlow or PyTorch used to create your model.
network-ID or default	If you know your network ID, use that, otherwise enter default.

3. Go to Compute Engine > VM instances. Find the instance with your VM name, and click SSH to connect to it.
4. From your VM, create an environment variable containing the name of your TPU:

   $ export TPU_NAME=tpu-name
   

Setting up a Compute Engine VM and Cloud TPU resources

You can allocate and start your VM and TPU resources using the ctpu utility, gcloud commands, or the Cloud Console.

ctpu

Run the following command in the Cloud shell. The ctpu utility creates the Compute Engine VM and Cloud TPU resources together and gives them the same name.

$ ctpu up [optional: --name --zone --tpu-size --machine-type --disk-size-gb]

Parameter	Description
name	Specifies the name for both the Compute Engine VM and the Cloud TPU.
zone	The zone where you plan to create your Cloud TPU. For example, us-central1-b.
tpu-size	This is the TPU type to use. The default is v2-8. See Types and zones for the supported TPU types and zones.
machine type	Specifies the machine type to use for your Compute Engine VM. Select a machine type from the drop down menu.
disk-size-gb	The disk size in GB for the VM. For example, 300GB. If a size is not specified, it defaults to 20GB.

gcloud commands

The Cloud SDK is a set of tools that you can use to interact with Google Cloud in the Cloud shell.

1. Install the gcloud command-line tool via the Cloud SDK.
2. Use the gcloud command-line tool to specify your Google Cloud project:

   $ gcloud config set project project-name
   

3. Specify the zone where you plan to create your Compute Engine VM and Cloud TPU resource. For this example, use the us-central1-b zone:

   $ gcloud config set compute/zone zone
   

   Cloud TPU is available in the following zones:

   US

   TPU type (v2)	TPU v2 cores	Total TPU memory	Region/Zone
   v2-8	8	64 GiB	us-central1-b us-central1-c us-central1-f
   v2-32	32	256 GiB	us-central1-a
   v2-128	128	1 TiB	us-central1-a
   v2-256	256	2 TiB	us-central1-a
   v2-512	512	4 TiB	us-central1-a
   TPU type (v3)	TPU v3 cores	Total TPU memory	Available zones
   v3-8	8	128 GiB	us-central1-a us-central1-b us-central1-f

   Europe

   TPU type (v2)	TPU v2 cores	Total TPU memory	Region/Zone
   v2-8	8	64 GiB	europe-west4-a
   v2-32	32	256 GiB	europe-west4-a
   v2-128	128	1 TiB	europe-west4-a
   v2-256	256	2 TiB	europe-west4-a
   v2-512	512	4 TiB	europe-west4-a
   TPU type (v3)	TPU v3 cores	Total TPU memory	Available zones
   v3-8	8	128 GiB	europe-west4-a
   v3-32	32	512 GiB	europe-west4-a
   v3-64	64	1 TiB	europe-west4-a
   v3-128	128	2 TiB	europe-west4-a
   v3-256	256	4 TiB	europe-west4-a
   v3-512	512	8 TiB	europe-west4-a
   v3-1024	1024	16 TiB	europe-west4-a
   v3-2048	2048	32 TiB	europe-west4-a

   Asia Pacific

   TPU type (v2)	TPU v2 cores	Total TPU memory	Region/Zone
   v2-8	8	64 GiB	asia-east1-c

4. Create a Compute Engine VM to interact with your Cloud TPU.

   Since you specified the zone in the previous command, the VM instance is created in that zone.

   $ gcloud compute instances create vm-name\
      --machine-type=n1-standard-1 \
      --image-project=ml-images \
      --image-family=tensorflow-vm-image \
      --boot-disk-size=boot-disk-size \
      --scopes=cloud-platform
   

   Parameter	Description
   vm-name	Specify a name for the Compute Engine VM.
   machine-type	This is the type of machine to use for the VM. See machine types for the supported machine types.
   image-project	The project against which all image and image family references will be resolved. Use ml-images.
   image-family	This is the TensorFlow family of the image that the boot disk will be initialized with, for example, tf-1-15 or tf-2-1.
   boot-disk-size	The disk size in GB for the VM. For example, 300GB. If a size is not specified, it defaults to 20GB.
   scopes	Use cloud-platform.

   This will generate output similar to the following:

   NAME         ZONE           MACHINE_TYPE    PREEMPTIBLE INTERNAL_IP  EXTERNAL_IP    STATUS
   demo-vm-tpu  us-central1-b  n1-standard-1               10.138.0.2   35.247.15.162  RUNNING
   

5. Create a new Cloud TPU resource.
   Note: Billing begins as soon as the Cloud TPU is created, and continues until the time it is deleted. Check the Cloud TPU pricing page to estimate your costs. If you are using a dataset that requires a substantial download and processing phase, hold off on running this command until you are ready to use the Cloud TPU.

   $ gcloud compute tpus create tpu-name \
         --zone=zone \
         --network=network-id or default \
         --accelerator-type=tpu-type \
         --version=2.1
   

   Parameter	Description
   tpu-name	Specifies the name of the Cloud TPU. Use the same name you used for the Compute Engine VM name.
   zone	The zone where you plan to create your Cloud TPU. For example, us-central1-b.
   network	If you know your network ID, use that, otherwise enter default.
   accelerator-type	This is your TPU type. See TPU types for the supported TPU types for your zone.
   version	The TensorFlow version to use with your Cloud TPU. This must be the same version of TensorFlow used to create you model.

   This will generate output similar to the following:

   NAME         ZONE           ACCELERATOR_TYPE NETWORK_ENDPOINT  NETWORK  RANGE         STATUS
   demo-vm-tpu  us-central1-b  v2-8             10.240.1.2:8470   default  10.240.1.0/29 READY
   
   

6. Remotely connect to your Compute Engine VM:

   $ gcloud compute ssh vm-name --zone=zone
   

7. Create an environment variable containing the name of your TPU:

   $ export TPU_NAME=tpu-name
   

Console

Note: Use the same name and zone for the VM and TPU resources.
1. Create and start your VM.
   1. Go to Compute Engine > VM instances on the left-hand navigation bar and click CREATE INSTANCE.
   2. On the Create an instance page specify an instance name, the region, and the machine type.

   Parameter	Description
   name	Specifies the name of the Compute Engine VM. You can specify any instance name, but use the same one for both the VM instance and the Cloud TPU.
   region	If you have set up a Cloud Storage bucket, this should match the Location of the bucket.
   machine type	Specifies the machine type to use for your Compute Engine VM. Select a machine type from the drop down menu.

2. Create, start and connect to your Cloud TPU.
   Note: Cloud TPU charges begin when you create your Cloud TPU. If you need to download a dataset, do that after you set up your VM, but before you set up your Cloud TPU to avoid Cloud TPU charges while the download proceeds.
1. Go to Compute Engine > TPUs on the left-hand navigation bar and click CREATE TPU NODE.
2. On the Create a Cloud TPU page use the menu pulldowns to specify the TPU name, the zone, TPU type, TPU software version, network, and an internal IP address for the Cloud TPU to use.

Parameter	Description
name	Specifies the name of the Cloud TPU. Use the same name you used for the Compute Engine VM name.
zone	The zone where you plan to create your Cloud TPU. For example, us-central1-b.
TPU type	This is your TPU type. See TPU types for the supported TPU types for your zone.
TPU software version	The TensorFlow or PyTorch version to use with your Cloud TPU. This must be the same version of TensorFlow or PyTorch used to create your model.
network-ID or default	If you know your network ID, use that, otherwise enter default.

3. Go to Compute Engine > VM instances. Find the instance with your VM name, and click SSH to connect to it.

Stopping your Cloud TPU resources

Charges for Cloud TPU resources begin when the Cloud TPU starts, even if it is not yet actively training a model. To avoid being charged while the Cloud TPU is inactive, you can stop it and restart it when you are ready to train a model.

Note: You can also start only the Compute Engine VM to download the dataset and start the Cloud TPU later when you are ready to start training. See the --vm-only and --tpu-only flags under ctpu up in the CTPU reference.

This section shows how to stop the Cloud TPU using the ctpu utility, gcloud commands, or the Cloud Console.

ctpu

1. Run the ctpu status command, specifying the zone where your Cloud TPU is set up.

   $ ctpu status --zone=zone

   This will display the status of Compute Engine VM and Cloud TPU resources within the zone.

   Your cluster is running!
   Compute Engine VM:  RUNNING
   Cloud TPU:          RUNNING
   

2. If the Cloud TPU resource is running, use the following command to stop it.

   $ ctpu pause --zone=zone

   This stops the Compute Engine VM and deletes the Cloud TPU resources in the specified zone. To only restart the Compute Engine VM, run the following command, including any other flags you need for the VM:

   $ ctpu up --vm-only --zone=zone 

   To restart the Cloud TPU, run ctpu up with the flags you set when you first started the Cloud TPU.

gcloud

1. Run the following command in the Cloud shell to list the available Cloud TPU resources in your zone.

   $ gcloud compute tpus list --zone=zone
   

   The Cloud TPU resource in your zone is displayed:

   NAME       ZONE           ACCELERATOR_TYPE  NETWORK_ENDPOINT   NETWORK  RANGE          STATUS
   demo-tpu   us-central1-b  v2-8              10.240.1.2:8470    default  10.240.1.0/29  READY
   

2. Run the following command to stop the Cloud TPU:

   $ gcloud compute tpus stop tpu-name --zone=zone
   

Console

Select Compute Engine > TPUs from the left-hand navigation bar. Click STOP from the menu bar at the top of the page.

To restart the Cloud TPU, click START.

Viewing your Compute Engine VM and Cloud TPU resources

This section shows how you can view your currently active VM and TPU resources using the ctpu utility, gcloud commands, or the Cloud Console.

ctpu

Run the ctpu status command and specify the zone where your Compute Engine VM and Cloud TPU resources are set up.

$ ctpu status --zone=zone 

This will display the status of Compute Engine VM and Cloud TPU resources within the zone.

Your cluster is running!
Compute Engine VM:  RUNNING
Cloud TPU:          RUNNING

If no resources are currently set up, the output will just show dashes for the VM and TPU. If one resource is active and the other is not, you will see a message saying the status is unhealthy. You need to start or restart whichever resource is not running.

gcloud

1. Run the following command in the Cloud shell to list the available Compute Engine VM resources in specific zones. In this example, VM resources in us-central1-b and europe-west4-a will be displayed:

   $ gcloud compute instances list --filter="zone:( us-central1-b europe-west4-a )"
   

   The above command prints the details of the VM resources you've created. For example:

   NAME      ZONE           MACHINE_TYPE   PREEMPTIBLE  INTERNAL_IP  EXTERNAL_IP     STATUS
   demo-tpu  us-central1-b  n1-standard-1               10.128.0.33  35.232.214.205  RUNNING
   
   

2. Run the following command from the Cloud shell to list the available Cloud TPU resources in your zone. In this example we show the selected zone to be us-central1-b.

   $ gcloud compute tpus list --zone=us-central1-b
   

   The Cloud TPU resource in us-central1-b is displayed:

   NAME       ZONE           ACCELERATOR_TYPE  NETWORK_ENDPOINT   NETWORK  RANGE          STATUS
   demo-tpu   us-central1-b  v2-8              10.240.1.2:8470    default  10.240.1.0/29  READY
   

Console

1. From the left navigation menu, select Compute Engine > TPUs.

   A list of all active Compute Engine resources appears.

2. From the left navigation menu, select Compute Engine > VM Instances.

   A list of all active Compute Engine resources appears.

Deleting your Compute Engine VM and Cloud TPU resources

You can delete your VM and TPU resources using the ctpu utility, gcloud commands, or the Cloud Console.

ctpu

Run the following command from the Cloud shell. The ctpu utility deletes the Compute Engine VM and Cloud TPU resources together.

Note: Be sure to specify the zone if you have changed it from the default. You do not need to specify other optional flags to delete the resources.

$ ctpu delete [optional: --zone]

gcloud

Run the following command from the Cloud shell to delete your Cloud TPU and Compute Engine VM resources. You must include the Cloud TPU resource name, the zone, and the project ID on the command line. Delete the Cloud TPU first since you need a running Compute Engine VM to delete the Cloud TPU resource.

1. Delete your Cloud TPU resource:

   $ gcloud compute tpus delete tpu-name --project=project-name --zone=zone
   

2. Delete your Compute Engine instance:

   $ gcloud compute instances delete vm-name --project=project-name --zone=zone
   

3. Delete the VPC network that Google automatically created as part of the Cloud TPU setup.

   Go to the VPC Networking page on the Cloud Console.

4. Select the VPC network. The network entry starts with cp-to-tp-peering in the ID.

   At the top of the page, click Delete to delete the selected VPC network.

5. When you've finished deleting the resources, use the gsutil command to delete any Cloud Storage buckets you created. Replace bucket-name with the name of your Cloud Storage bucket:

   $ gsutil rm -r gs://bucket-name
   

Console

1. Delete your VM.
   1. Go to Compute Engine > VM instances from the left-hand navigation
   bar.
   2. Select your VM instance from the list. Click the trash can icon at the top of the page.
2. Delete your Cloud TPU.
   1. Go to Compute Engine > TPUs on the left-hand navigation bar.
   2. Select your TPU resource from the list. Click the trash can icon at the top of the page.