Skip to navigation

Create a GPU VM

Creating a GPU VM gives you a Linux server with an NVIDIA GPU attached, ready for AI, rendering or video work. The portal’s Deploy page is a single form with a running cost summary and the Deploy button in the bottom bar. It works like the Cloud VM Deploy page, with GPU plans and GPU-compatible images.

Which options to choose

ChoiceRecommendedWhy
PlanThe smallest GPU and memory that fits your modelGPU plans are billed for every hour the VM exists, running or stopped
ImageA GPU-compatible Ubuntu templateMost AI tools and NVIDIA’s own packages target Ubuntu
AccessAn SSH keySafer than a password
NetworkPublic IPSo you can SSH in and download models. Use Private + NAT for workers that only fetch jobs and data, and never accept connections.

How it works

  1. You choose a location, a GPU plan, a Linux image, SSH keys and the network, and click Deploy.
  2. The portal checks your billing state and wallet balance, then queues the VM. It shows Provisioning, then Running.
  3. Billing starts when creation finishes. A create that fails isn’t billed.
  4. You sign in, run nvidia-smi, and install the NVIDIA driver if the image doesn’t include it.

Before you begin

  • An IBEE account with a verified identity, and write access to the workspace.
  • Credit in your wallet for the GPU plan’s hourly price (Billing and usage).
  • An SSH key. You can also create one on the Deploy page.
  • GPUs are limited per location. The Deploy page only lets you pick as many VMs as there are free GPUs.

Create a GPU VM in the portal

1

Open the Deploy page

In the portal sidebar, open Infrastructure → GPU VMs, then click Deploy GPU VM.

2

Select a location

Under Location, pick a region tab and a data center. GPU models differ between locations.

3

Choose a GPU plan

Under Choose Plan, select a plan. Each row lists the GPU model and its Memory, plus vCPU, RAM, Disk, and the price per month and per hour. GPU VMs are billed Hourly; the monthly figure is the hourly rate × 731 hours.

4

Select the operating system

Under Operating System → Templates, pick a family card, then a version from OS Version. Only GPU-compatible images are listed. You can also start from Snapshots or Backups of another GPU VM. ISOs aren’t available: the tab shows “No ISOs available for this account.”

5

Add SSH keys

Open Select SSH Keys and pick one or more saved keys, or click Add Key to create or import one. Without a key, the VM gets a generated password, shown under Overview → Connection Details.

6

Configure the network

Under Network Configuration, choose Public IP (default), Private + NAT or Private only, and a VPC and subnet if needed. Under Firewall, pick a group or leave it empty to attach the default group, which allows SSH (22), HTTP (80), HTTPS (443), RDP (3389) and ports 8000–8080. See Networking for VMs.

7

Set VM details

Under VM Details, set Instances (up to 5, fewer when free GPUs or public addresses are short, and 1 with a specific Reserved IP) and a Hostname (lowercase letters, numbers and hyphens). With several instances, each gets <base>-1, <base>-2 and so on.

8

Deploy

Check the price in the bottom bar and click Deploy (or Deploy N VMs). If the billing check passes, you return to the GPU Virtual Machines list and the VM shows Provisioning. If not, a Billing needs attention popup appears: click Add Credits, top up, and deploy again.

Create a GPU VM with the CLI

GPU VMs use ibee gpus (Cloud VMs use ibee vms). Export an API token with the vm.read and vm.write scopes and your workspace ID first:

export IBEE_TOKEN="ibee_prod_key_xxxxxxxxxxxx"
export IBEE_WORKSPACE_ID="710995"
ibee compute sites
ibee compute plans --vm-type gpu --site-id SITE_ID
ibee compute images --vm-type gpu --site-id SITE_ID
ibee gpus create --name ml-01 \
--site-id SITE_ID \
--plan-id GPU_PLAN_ID \
--template-id GPU_TEMPLATE_ID \
--os-type linux --os-distro ubuntu \
--cpu CPU --ram-mb RAM_MB --disk-gb DISK_GB \
--gpu-model GPU_MODEL --gpu-count GPU_COUNT \
--billing-interval HOURLY \
--ssh-keys "$(cat ~/.ssh/id_ed25519.pub)"

The CLI sends only what you pass: copy the vCPUs, memory, disk, GPU model and GPU count from the plan, and the OS from the image. --billing-interval HOURLY (or MONTHLY) makes the CLI read the plan and send its billing selection, as the portal does. The command prints an operation_id; run ibee ops get OPERATION_ID until status is succeeded or failed. To check capacity first, call GET /compute/plans?vm_type=gpu&site_id=SITE_ID&include_availability=true&workspace_id=...: each plan’s availability is available, unavailable or unknown. The check is advisory and doesn’t hold a GPU for you.

The API operation is POST /compute/gpu-vms, with an X-Idempotency-Key header. It takes the same fields as a Cloud VM create plus gpu_model and gpu_count, and accepts Linux images only (os_type: linux).

Verify

When the VM shows Running, sign in with the command from Overview → Connection Details:

ssh ubuntu@203.0.113.20
nvidia-smi

nvidia-smi prints a table with the GPU model, its memory, the driver version and the CUDA version the driver supports. If you see nvidia-smi: command not found, install the driver as below.

Install the NVIDIA driver

Some images come without an NVIDIA driver. On Ubuntu, install the recommended driver and restart:

sudo apt update
sudo apt install -y ubuntu-drivers-common
sudo ubuntu-drivers install
sudo reboot

After the restart, sign in again and run nvidia-smi. Install the CUDA toolkit or a framework such as PyTorch only if your software needs it; GPU monitoring needs only the driver.

Manage your GPU VM

The VM’s page has Actions (Start, Stop, Reboot, Delete; Console shows Soon) and these tabs:

TabWhat it’s for
OverviewSpecs, network and Connection Details
MonitoringCPU, memory, disk and network charts. See Monitoring.
GPU MetricsGPU utilization, GPU memory, temperature and power. Click Enable monitoring first. See GPU Metrics.
Backups / SnapshotsRecovery points. See Backups and Snapshots.
SettingsChange Plan (vCPU, memory and disk; the GPU stays the same), IPv4 & Reserved IP, VPC Networks and Access. Change OS and Hostname aren’t available yet.

A stopped GPU VM is billed at the full hourly rate. Delete it when you’ve finished, and take a snapshot first if you want to come back to it.

Limits

Value
VMs per deploy1 to 5, limited by free GPUs in the location
Operating systemsLinux only
Billing in the portalHourly, from when creation finishes until the VM is deleted
Change the GPU modelNot possible in place; create a new GPU VM
CLI --timeout 3600Waits up to 3600 seconds (1 hour) for the create to finish

Troubleshooting

You seeWhat it meansWhat to do
Please select a VM plan, Please select an Operating System version or Each instance needs a unique hostnameA required field is empty or invalidFix the named field and click Deploy again.
Billing needs attentionYour billing state or wallet balance blocks the deployClick Add Credits, top up, then deploy again.
Only 1 instance can be chosenFew free GPUs (or public addresses) in the location, or you picked a specific Reserved IPDeploy one, or try another plan or location.
nvidia-smi: command not foundThe image has no NVIDIA driverInstall the driver.
NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driverThe driver is installed but not loadedRun sudo reboot. If it persists, reinstall the driver.
The VM stays in ProvisioningCreation is still running or has stalledCheck Activity logs. If nothing changes, contact IBEE support.