Creating a VM with a GPU

If you don't have the Yandex Cloud command line interface yet, install and initialize it.

The folder specified in the CLI profile is used by default. You can specify a different folder using the --folder-name or --folder-id parameter.

1. View a description of the CLI create VM command:

   yc compute instance create --help
   

2. Prepare the key pair (public and private keys) for SSH access to the VM.

3. Select a public image.

   To get a list of available images, run the following command:

   yc compute image list --folder-id standard-images
   +----------------------+----------------------------------+-----------------------+----------------------+--------+
   |          ID          |               NAME               |         FAMILY        |     PRODUCT IDS      | STATUS |
   +----------------------+----------------------------------+-----------------------+----------------------+--------+
   ...
   | fdv7ooobjfl3ts9gqp0q | windows-2016-gvlk-gpu-1548913814 | windows-2016-gvlk-gpu | dqnnc72gj2ist3ktjj1p | READY  |
   | fdv4f5kv5cvf3ohu4flt | ubuntu-1604-lts-gpu-1549457823   | ubuntu-1604-lts-gpu   | dqnnb6dc7640c5i968ro | READY  |
   ...
   +----------------------+----------------------------------+-----------------------+----------------------+--------+
   

   For VM instances running on the Intel Broadwell with NVIDIA® Tesla® V100 and Intel Cascade Lake with NVIDIA® Tesla® V100 platforms, a special Ubuntu images is available in Cloud Marketplace: Windows — 2016 Datacenter GPU (windows-2016-gvlk-gpu), Ubuntu — 16.04 LTS GPU (ubuntu-1604-lts-gpu) and 20.04 LTS GPU (ubuntu-2004-lts-gpu). The images have NVIDIA drivers pre-installed.

   For VM instances running on the Intel Ice Lake with NVIDIA® Tesla® T4 platform, an Ubuntu image is available: 20.04 LTS GPU A100 (ubuntu-2004-lts-gpu).

   For VM instances running on the AMD EPYC™ with NVIDIA® Ampere® A100 platform, a special Ubuntu image is available: 20.04 LTS GPU A100 (ubuntu-2004-lts-gpu-a100). We recommend using a standard image from Yandex Cloud. You can also install the drivers on another standard image yourself or create a custom image with preinstalled drivers.

4. Create a VM in the default folder:

   yc compute instance create \
     --name gpu-instance \
     --zone ru-central1-a \
     --platform=gpu-standard-v3 \
     --cores=8 \
     --memory=96 \
     --gpus=1 \
     --network-interface subnet-name=default-ru-central1-a,nat-ip-version=ipv4 \
     --create-boot-disk image-folder-id=standard-images,image-family=ubuntu-1604-lts-gpu \
     --ssh-key ~/.ssh/id_rsa.pub
   

   Where:

   • name: VM name.

     Note

     The VM name is used to generate an internal FQDN only once: when creating a VM. If the internal FQDN is important to you, choose an appropriate name for the VM at the creation stage.

   • zone: Availability zone.

   • platform: Platform ID:

     • gpu-standard-v1 for Intel Broadwell with NVIDIA® Tesla® V100.
     • gpu-standard-v2 for Intel Cascade Lake with NVIDIA® Tesla® V100.
     • gpu-standard-v3 for AMD EPYC™ with NVIDIA® Ampere® A100.

   • cores: Number of vCPUs.

   • memory: RAM.

   • gpus: Number of GPUs.

   • preemptible: If required, make your VM preemptible.

   • create-boot-disk: Image of the OS. ubuntu-1604-lts-gpu: Ubuntu 16.04 LTS GPU with CUDA drivers.

   • nat-ip-version: Public IP. To create a VM without a public IP address, disable the nat-ip-version=ipv4 option.

   This creates a VM named gpu-instance with a single GPU, 8 vCPUs, and 96 GB RAM:

   yc compute instance get --full gpu-instance
   

   Result:

   name: gpu-instance
   zone_id: ru-central1-a
   platform_id: gpu-standard-v3
   resources:
     memory: "103079215104"
     cores: "8"
     core_fraction: "100"
     gpus: "1"
   status: RUNNING
   ...
   

To create a VM, use the Create method for the Instance resource.

If you don't have Terraform, install it and configure the Yandex Cloud provider.

1. In the configuration file, describe the parameters of resources that you want to create:

   resource "yandex_compute_instance" "vm-1" {
   
     name        = "vm-with-gpu"
     platform_id = "gpu-standard-v3"
   
     resources {
       cores  = <number of vCPU cores>
       memory = <RAM amount, GB>
       gpus   = <number of GPUs>
     }
   
     boot_disk {
       initialize_params {
         image_id = "fdv4f5kv5cvf3ohu4flt"
       }
     }
   
     network_interface {
       subnet_id = "${yandex_vpc_subnet.subnet-1.id}"
       nat       = true
     }
   
     metadata = {
       ssh-keys = "<username>:<SSH key contents>"
     }
   }
   
   resource "yandex_vpc_network" "network-1" {
     name = "network1"
   }
   
   resource "yandex_vpc_subnet" "subnet-1" {
     name       = "subnet1"
     zone       = "<availability zone>"
     network_id = "${yandex_vpc_network.network-1.id}"
   }
   

   Where:

   • yandex_compute_instance: Description of the VM:

     • name: VM name.

     • platform_id: ID of the platform:

       • gpu-standard-v1 for Intel Broadwell with NVIDIA® Tesla® V100.
       • gpu-standard-v2 for Intel Cascade Lake with NVIDIA® Tesla® V100.
       • gpu-standard-v3 for AMD EPYC 7662 with NVIDIA® Ampere® A100.

     • resources: The number of vCPU cores and the amount of RAM available to the VM. The values must match the selected platform.

     • boot_disk: Boot disk settings. Specify the ID of the selected image. You can get the image ID from the list of public images.

       For VM instances running on the Intel Broadwell with NVIDIA® Tesla® V100 and Intel Cascade Lake with NVIDIA® Tesla® V100 platforms, a special Ubuntu images is available in Cloud Marketplace: Windows — 2016 Datacenter GPU (windows-2016-gvlk-gpu), Ubuntu — 16.04 LTS GPU (ubuntu-1604-lts-gpu) and 20.04 LTS GPU (ubuntu-2004-lts-gpu). The images have NVIDIA drivers pre-installed.

       For VM instances running on the Intel Ice Lake with NVIDIA® Tesla® T4 platform, an Ubuntu image is available: 20.04 LTS GPU A100 (ubuntu-2004-lts-gpu).

       For VM instances running on the AMD EPYC™ with NVIDIA® Ampere® A100 platform, a special Ubuntu image is available: 20.04 LTS GPU A100 (ubuntu-2004-lts-gpu-a100). We recommend using a standard image from Yandex Cloud. You can also install the drivers on another standard image yourself or create a custom image with preinstalled drivers.

     • network_interface: Network settings. Specify the ID of the selected subnet. To automatically assign a public IP address to the VM, set nat = true.

     • metadata: In the metadata, pass the public key for VM access via SSH. For more information, see VM instance metadata.

   • yandex_vpc_network: Description of the cloud network.
   • yandex_vpc_subnet: Description of the subnet your VM will connect to.

   Note

   If you already have suitable resources, such as a cloud network and subnet, you don't need to describe them again. Use their names and IDs in the appropriate parameters.

   For more information about resources that you can create with Terraform, please see the provider documentation.

2. Make sure that the configuration files are correct.

   1. In the command line, go to the directory where you created the configuration file.

   2. Run the check using the command:

      terraform plan
      

   If the configuration is described correctly, the terminal displays a list of created resources and their parameters. If there are errors in the configuration, Terraform points them out.

3. Deploy the cloud resources.

   1. If the configuration doesn't contain any errors, run the command:

      terraform apply
      

   2. Confirm that you want to create the resources.

   Afterwards, all the necessary resources are created in the specified folder. You can check that the resources are there with the correct settings using the management console.