Skip to navigation

GPU VMs

A GPU VM is a virtual machine with an NVIDIA graphics processor (GPU) attached to it. GPUs do thousands of calculations at once, which makes them much faster than ordinary processors for AI training and inference, video encoding and 3D rendering. Apart from the GPU, a GPU VM works like a Cloud VM: you create it in the portal, CLI or API, sign in with SSH, and pay by the hour.

Ready to create one? Follow Create a GPU VM.

When to use a GPU VM

You want to…Use
Train or fine-tune AI models, serve models (inference), render, or encode videoGPU VM (this page)
Run websites, APIs, databases or ordinary appsCloud VM: cheaper when you don’t need a GPU
Have a whole physical server to yourselfBare metal server

How it works

  1. You pick a GPU plan. Each plan combines a GPU model and its memory with vCPUs, RAM and a root disk.
  2. You pick a GPU-compatible Linux image. Only images that work with GPUs are listed on the GPU Deploy page.
  3. IBEE creates the VM with the GPU passed through to it. Billing starts when creation finishes.
  4. You sign in, check the GPU with nvidia-smi, and install the NVIDIA driver if the image doesn’t include one.

Before you begin

  • An IBEE account with a verified identity, and write access to the workspace.
  • Credit in your wallet. GPU plans cost more than Cloud VM plans; the Deploy page shows each plan’s hourly and monthly price.
  • An SSH key. GPU VMs run Linux.

Key facts

QuestionAnswer
Are NVIDIA drivers or CUDA installed?It depends on the image. Run nvidia-smi after you sign in. If it isn’t found, install a compatible NVIDIA driver inside the VM. CUDA is only needed by your own software. See Create a GPU VM.
How is it billed?Hourly in the portal, for every hour the VM exists.
Am I billed while it’s stopped?Yes, at the same hourly rate as when it’s running, until you delete it. Delete the VM to stop the charge.
Can I change the GPU?No. Settings → Change Plan changes vCPUs, memory and disk, but the GPU model stays the same. To use another GPU, take a snapshot and create a new GPU VM with the GPU you want.
Which disk keeps my data?The root disk and any attached volumes. Both keep their data through stop, start, reboot and resize. There’s no temporary scratch disk that is wiped.
How many at once?Up to 5 per deploy, fewer when the location has fewer free GPUs.
Which operating systems?Linux only. The API refuses Windows for GPU VMs.
Is there a browser console?Not yet. The Console action shows Soon. Use SSH.
How fine is the billing?Usage is collected every 15 minutes and charged for the time the VM exists; the monthly estimate is the hourly rate × 731 hours.

Find GPU VMs in the portal

In the portal sidebar, open Infrastructure → GPU VMs. The GPU Virtual Machines list shows each VM’s Instance, OS, Location, IP Address and Status. Click Deploy GPU VM to create one, or click a row to open it.

How GPU VMs differ from Cloud VMs

GPU VMsCloud VMs
PlansGPU model and GPU memory, plus vCPU, RAM, disk and pricevCPU, memory, disk and included bandwidth. See Instance types.
Billing in the portalHourlyHourly, or Monthly where offered
ImagesGPU-compatible Linux imagesLinux and Windows images
Tabs on the VM pageOverview, Monitoring, GPU Metrics, Backups, Snapshots, SettingsOverview, Monitoring, Backups, Snapshots, Settings
CLIibee gpus …ibee vms …
API/compute/gpu-vms/compute/cloud-vms

Everything else works the same way: locations, SSH keys, network options, firewall groups, power actions, resizing, backups and snapshots.

GPU Metrics

The GPU Metrics tab charts GPU Utilization, VRAM Utilization (GPU memory), Temperature and Power Draw over 30m, 1h, 6h, 24h or 7d. It’s off until you turn it on.

1

Start the VM and the guest agent

The VM must be Running, with qemu-guest-agent running inside it. Check with systemctl status qemu-guest-agent.

2

Turn monitoring on

On the GPU Metrics tab, click Enable monitoring. IBEE installs a small agent through the guest agent. “It opens no VM network port.”

3

Install the driver if asked

If the tab shows NVIDIA guest driver required, install an NVIDIA driver inside the VM. Monitoring starts by itself once nvidia-smi works. CUDA is optional.

The tab also shows the driver version, CUDA version and agent version. Charts keep their history while the VM is stopped.

Troubleshooting

You seeWhat it meansWhat to do
nvidia-smi: command not foundThe image has no NVIDIA driverInstall a driver: see Create a GPU VM.
NVIDIA guest driver required on GPU MetricsThe monitoring agent is ready but nvidia-smi doesn’t workInstall the driver; monitoring then starts automatically.
QEMU Guest Agent is unreachableThe guest agent isn’t running in the VMRun sudo systemctl enable --now qemu-guest-agent, then click Repair monitoring.
GPU monitoring agent is offline”No heartbeat has arrived for more than five minutes.”Click Repair monitoring.
VM is stopped on GPU MetricsLive collection pauses while the VM is offStart the VM.
You can’t pick more GPU VMs on the Deploy pageThe location has no more free GPUs of that modelDeploy fewer, or pick another plan or location.