---
title: "Deploying GPU instances"
slug: "gpu-vm-deployment"
source: "https://app.cloudpe.com/help/gpu-vm-deployment"
updated: "2026-08-27T05:22:14.342Z"
---

# Deploying GPU instances

## Overview

GPU Cloud gives you NVIDIA-accelerated virtual machines for training, inference, rendering and other HPC workloads. GPU instances behave like standard instances — same images, volumes, networking and console — with GPU-specific selection, stock checks and telemetry layered on top.

The feature is live and lives under **COMPUTE** → **GPU Cloud** in the dashboard. GPU types available in the catalog include NVIDIA L4, NVIDIA RTX A4000, NVIDIA RTX A5000, NVIDIA A100, NVIDIA H200, NVIDIA L40S and RTX Pro 6000, offered from the Mumbai DC2 Zone A and Mumbai DC2 Zone B regions.

## Before you start

- Identity verification: your organization must have completed KYC before any cloud resource can be created.
- Wallet balance: GPU instances are billable on creation; keep enough prepaid balance (or an eligible postpaid entitlement) in the wallet, otherwise the create request is refused by the billing gate.
- Permissions: you need `vms:read` to see the GPU Cloud list and `vms:create` to launch. Day-2 actions use `vms:power`, `vms:console` and `vms:update`.
- A project in the right region: the project you launch into must belong to the same region as the GPU you select. Create or switch to a project in Mumbai DC2 Zone A or Mumbai DC2 Zone B first.
- Stock: GPU capacity is finite. The create wizard shows live availability per GPU type; a type with no free cards in that region cannot be launched until capacity frees up.

## Steps

1. In the dashboard sidebar, open **COMPUTE** → **GPU Cloud**. The **GPU Virtual Machines** page lists your existing GPU instances; use **Status** and **Region** to filter.

   ![](/kb/compute/gpu-vm-deployment-01-list.png)

2. Click **Create GPU VM** to open **Create GPU Virtual Machine**.

3. On the **GPU** step, under **Select GPU Type**, pick the accelerator you want. Availability is shown per region.

   ![](/kb/compute/gpu-vm-deployment-02-create.png)

4. Click **Next** to move to **Flavor** and choose the vCPU/RAM shape paired with that GPU. Use **Back** at any point to change an earlier choice.

5. On the **Image** step, choose an operating system. Pre-built accelerated images such as `Ubuntu-24.04-GPU`, `Ubuntu-22.04-GPU`, `Ubuntu-20.04-GPU` and `Ubuntu 24.04 GPU - vLLM` ship with GPU drivers already in place, so you do not have to install them yourself.

6. On the **Billing** step, select the billing cycle for the instance.

7. On the **Configure** step, set the instance name, project, network, security groups, SSH key and boot volume (storage policy options are Eco-NVMe, Standard-NVMe and Pro-NVMe). Submit to launch.

8. Open the instance from the list to reach its detail page. Use **Refresh status** while it provisions, then **Console**, **Reboot** and **Reset Password** for day-2 operations. The tabs **Overview**, **Billing**, **Volumes**, **Networking**, **Logs**, **Settings**, **Metrics** and **Alerts** cover accelerator details, cost breakdown, attached volumes and snapshots, interfaces and floating IPs, boot log and action history, tags and shelve scheduling, telemetry, and alert rules.

   ![](/kb/compute/gpu-vm-deployment-03-detail.png)

9. Once the instance is reachable over SSH, confirm the accelerator is visible to the guest with `nvidia-smi`.

## API

There is no public endpoint for launching GPU instances; use the dashboard. If you need to add wallet balance before launching, create a payment order:

```bash
curl -X POST https://app.cloudpe.com/api/v1/payments/create-order \
  -H "Authorization: Bearer <API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "amount": "<amount>",
    "organization_id": "<organization_id>"
  }'
```

`amount` is required; `organization_id` is optional and may be null.

## Limits & billing

- GPU instances meter three independent streams: compute (the GPU flavor), storage (boot volume plus any attached volumes) and network (public/floating IP). Each stream bills hourly against the wallet.
- Stopping a GPU instance stops compute metering — storage keeps billing. The physical GPU is released while the instance is stopped and may not be immediately available again when demand is high.
- Resizing is restricted: you can only move to another flavor with the same GPU model, and only on hourly-billed instances. Instances on a monthly or yearly subscription must be switched to hourly billing first, and instances under a committed-use plan are price-locked for the term.
- A resize on a running instance reboots it; resizing a stopped instance keeps it stopped and applies the new specs on next start.
- Multi-GPU flavors need all cards free on a single host, so a resize can still fail placement even when the stock badge shows capacity — a failed resize is revertible.
- GPU telemetry on the **Metrics** tab (utilisation, VRAM usage, temperature) requires the guest monitoring agent with GPU support; installation instructions appear on that tab if the agent is missing.

## Troubleshooting

| Message | What it means | What to do |
|---|---|---|
| `Complete identity verification to create cloud resources.` | Your organization has not finished KYC. | Complete identity verification, then retry the launch. |
| `This organization is not eligible to provision yet, so no projects were created. Resolve the listed requirements (Entitlement & Onboarding tab) and retry.` | Onboarding or entitlement requirements are outstanding, so no project exists to launch into. | Clear the listed requirements and retry. |
| `Project belongs to a different region; cannot create resources here.` | The selected project is not in the region that hosts the chosen GPU. | Switch to (or create) a project in the GPU's region and launch again. |
| `Billing gate misconfigured for this resource.` | The create request could not be priced or gated. | Retry; if it persists, raise a support ticket with the instance name and region. |
| `Failed to create payment order` | A wallet top-up order could not be raised. | Retry the top-up; if it keeps failing, contact support before retrying the launch. |
| `Failed to create volume: {…}` | An additional volume could not be provisioned during launch. | The instance itself may still be created — check the **Volumes** tab and create/attach the volume manually. |

## FAQ

**Do I need to install NVIDIA drivers?**
No, if you launch from one of the pre-built GPU images. They already include drivers, so `nvidia-smi` works on first boot.

**Can I change the GPU model later?**
No. Resize is limited to flavors with the same GPU model, and cannot cross to or from non-GPU flavors. To change accelerator, launch a new instance.

**Does stopping the instance stop all charges?**
No. Compute metering stops, but storage for the boot volume and any attached volumes continues.

**Where should I keep datasets?**
On attached block volumes rather than the boot disk, so data survives rebuilds and can be moved between instances.

**Can I see GPU utilisation without SSH?**
Yes — the **Metrics** tab shows GPU utilisation, VRAM usage and temperature once the guest agent is installed, and **Alerts** lets you set thresholds on those series.

## Related

- [Creating and managing virtual machines](/help/create-manage-virtual-machines)
- [Block storage volumes](/help/block-storage-volumes)
- [AI inference endpoints](/help/ai-inference-endpoints)