Skip to content

GPU Instances

GPU (Graphics Processing Unit) instances provide hardware-accelerated computing for workloads that require massive parallel processing power. Unlike standard VMs, GPU instances have one or more physical NVIDIA GPUs attached via PCI passthrough, giving the guest OS direct access to the GPU hardware.

Typical use cases include:

  • Machine Learning / AI — model training, inference, and fine-tuning
  • Rendering — 3D rendering, video encoding, and GPU-accelerated visualization
  • High-Performance Computing (HPC) — scientific simulations, computational chemistry, and data analysis

GPU Flavors

GPU flavors are available in two regions with different GPU hardware. Use openstack flavor list to see all available flavors in your project.

NVIDIA L4 — Region 1 (ch-zh1)

The NVIDIA L4 is a versatile GPU based on the Ada Lovelace architecture, offering 24 GB GDDR6 memory. It is well-suited for inference, video processing, and light ML workloads.

Flavor vCPU RAM Disk GPU Disk Type
p1d.small 1 2 GB 50 GB 1× L4 Shared
p1d.medium 1 4 GB 80 GB 1× L4 Shared
p1d.large 2 8 GB 120 GB 1× L4 Shared
p1d.xlarge 4 16 GB 240 GB 1× L4 Shared
p1d.2xlarge 8 32 GB 480 GB 1× L4 Shared
p1d.4xlarge 16 64 GB 960 GB 1× L4 Shared
p1gd.4xlarge 16 64 GB 960 GB 1× L4 Dedicated

The p1d flavors share the disk between multiple tenants on the hypervisor, while p1gd flavors provide a dedicated disk for higher I/O performance.

openstack flavor list --long | grep p1d

NVIDIA RTX 6000 Pro Blackwell — Region 2 (ch-ge1)

The RTX 6000 Pro is based on the Blackwell architecture and offers 96 GB GDDR7 memory. It is ideal for large-model inference, rendering, and compute-intensive workloads.

Flavor Prefix GPU VRAM Region
a1g6000d.* 1× RTX 6000 Pro 96 GB GDDR7 ch-ge1 
openstack flavor list --long | grep a1g6000d

NVIDIA H200 — Region 2 (ch-ge1)

Coming Soon

The H200 flavors are not yet publicly available. They are listed here for planning purposes. Contact support if you need early access.

The NVIDIA H200 is based on the Hopper architecture and features 141 GB HBM3e memory. It is designed for the most demanding AI training and HPC workloads. H200 flavors support multi-GPU configurations with up to 8 GPUs per instance.

Flavor vCPU RAM Disk GPUs
a1g200d.16xlarge 48 192 GB 960 GB 1× H200
a1g200d.32xlarge 96 384 GB 960 GB 2× H200
a1g200d.64xlarge 192 768 GB 960 GB 4× H200
a1g200d.128xlarge 240 1536 GB 1920 GB 8× H200

GPU Images

GPU instances require a GPU-enabled image with pre-installed NVIDIA drivers. The following Ubuntu GPU images are available:

OS Versions GPU Variants
Ubuntu 22.04, 24.04 GPU, GPU UEFI, GPU UEFI CUDA
  • GPU — standard GPU image with NVIDIA drivers
  • GPU UEFI — UEFI-bootable GPU image with NVIDIA drivers
  • GPU UEFI CUDA — UEFI-bootable GPU image with NVIDIA drivers and CUDA toolkit pre-installed

Use the CUDA variant if you need the CUDA toolkit for ML frameworks. For inference-only workloads without CUDA dependencies, the standard GPU image is sufficient.

openstack image list --public | grep -i gpu

See Images for general image management.

Launch via Horizon

To deploy a GPU instance via the Horizon dashboard:

  1. Navigate to Compute > Instances and click Launch Instance.
  2. In the Details tab, enter a name and select an availability zone in the appropriate region (ch-zh1 for L4, ch-ge1 for RTX 6000 Pro / H200).
  3. In the Source tab, select a GPU image (e.g. Ubuntu 24.04 GPU UEFI CUDA) and create a new volume.
  4. In the Flavor tab, select a GPU flavor (e.g. p1d.xlarge or a1g6000d.*).
  5. Complete the remaining tabs (Networks, Security Groups, Key Pair) as usual.
  6. Click Launch Instance.

See Launch Instance for the full wizard walkthrough.

Launch via CLI

openstack server create gpu-vm \
    --image "Ubuntu 24.04 GPU UEFI CUDA" \
    --flavor p1d.xlarge \
    --key-name my-key \
    --network my-net \
    --security-group default \
    --availability-zone ch-zh1-az2

For RTX 6000 Pro instances in Region 2:

openstack server create gpu-vm-rtx \
    --image "Ubuntu 24.04 GPU UEFI CUDA" \
    --flavor a1g6000d.4xlarge \
    --key-name my-key \
    --network my-net \
    --security-group default \
    --availability-zone ch-ge1-az1

Tip

Use openstack flavor list | grep -E 'p1d|a1g6000d|a1g200d' to find all available GPU flavors. A complete flavor overview is also available on our website: Flavor Overview.

Verify the GPU

After the instance is active, SSH into it and verify that the GPU is visible:

# List all NVIDIA GPUs
nvidia-smi -L

# Show full GPU status (memory, processes, temperature)
nvidia-smi

# Check CUDA toolkit version (GPU UEFI CUDA images only)
nvcc --version

Example output for an L4 instance:

$ nvidia-smi -L
GPU 0: NVIDIA L4 (UUID: GPU-xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx)

If the GPU is not visible, ensure you used a GPU image and a GPU flavor. Check the device nodes:

ls -la /dev/nvidia*

Notes & Limitations

  • Region availability — L4 flavors (p1d.*, p1gd.*) are only available in Region 1 (ch-zh1). RTX 6000 Pro (a1g6000d.*) and H200 (a1g200d.*) are only available in Region 2 (ch-ge1).
  • No live migration — GPU instances cannot be live-migrated. If the underlying host requires maintenance, the instance must be stopped and restarted.
  • Quota — GPU flavors consume GPU quota. Check your quota with openstack quota show. If you need additional GPU quota, contact support.
  • Image requirement — Only GPU-enabled images include the NVIDIA drivers. Booting a GPU flavor with a standard (non-GPU) image will not make the GPU accessible.
  • H200 availability — The a1g200d.* flavors are not yet publicly available. Contact support for early access.