Overview

The powerful Graphics Processing Unit (GPU) VMs excel at massive parallel processing. This specialized hardware best accelerates compute-intensive workloads, such as training complex AI models, machine learning inference, and high-speed 3D rendering.

Cloud GPU VMs come with predefined templates. You cannot create new templates or modify the existing templates. Use Cloud API standard operations with template read access and volume management to deploy Cloud GPU VMs. Plan your storage requirements carefully and select the best template for your use case; otherwise, you will have to use additional volumes for your storage needs.

Cloud GPU VM configuration

Dedicated instance specifications and configuration templates

The server sizing model allocates Cloud GPU VMs with corresponding dedicated CPU cores and RAM based on available host capacity. The architecture utilizes PCIe passthrough for direct hardware access and optimal performance.

Template specifications

You may choose between the following four template sizes. The templates can only be used with the Cloud GPU VMs. CPU and RAM allocate proportionally to the number of GPUs. Resources utilize dedicated cores with limited flexibility.

Warning: Configuration templates are created during provisioning and cannot be changed later.

The breakdown of resources is as follows:

Template	GPU Model	GPU Type	Number of GPUs	Dedicated CPUs	RAM (GiB)	NVMe Storage (GB)
S	NVIDIA H200	H200 PCIe	1	15	267	1024
M	NVIDIA H200	H200 PCIe	2	30	534	1536
L	NVIDIA H200	H200 PCIe	4	60	1068	2048
XL	NVIDIA H200	H200 PCIe	8	127	2136	4096

Dedicated resource model

• CPU Cores: Dedicated AMD EPYC Turin (non-shared) allocation, with a fixed ratio proportional to the number of GPUs.

• Memory: Fixed ratio based on host specifications.

• Flexibility: Static resource allocation without dynamic scaling. For more information, see Known Constraints.

• Counters: The use of Cloud GPU VMs' vCPU and RAM counts into existing Virtual Data Center (VDC) resource usage. However, dedicated resource usage counters are enabled for Cloud GPU VMs. These counters permit granular monitoring of vCPUs, which differ from Dedicated Core Servers for the enterprise VM instances and SSD block storage.

Storage architecture and planning

Provision boot volumes

The first connected volume serves as the boot volume, containing the operating system and required system files. Provision boot volumes with adequate capacity at the initial Cloud GPU VM provisioning, because they use fixed sizing and cannot be detached or upscaled after deployment. Any storage device, including the CD-ROM, can be selected as the boot volume. You may also boot from the network.

Adding additional volumes

Cloud GPU VM storage separates boot volumes from data volumes. You can use additional volumes for datasets requiring expansion; these volumes scale independently to accommodate growing storage requirements without interrupting operations and can be managed independently.

Add-on Block Storage

Additional storage volumes attached after server creation provide scalable capacity.

• Included storage: A default Cloud GPU VM comes ready with high-speed direct-attached NVMe storage.

• External storage: You may attach up to 23 additional volumes of HDD or SSD (Standard or Premium) block storage. Added HDD and SSD devices, as well as CD-ROMs, can be unmounted and deleted any time after the Cloud GPU VM is provisioned for use. For more information, see Set Up Block Storage.

• Flexibility: These additional volumes support scaling up, detaching, and attaching independently without affecting the boot volume.

Create and use images and snapshots

Images and snapshots can be created from and copied to direct-attached storage, block storage devices, and CD-ROM drives. Also, direct-attached storage volume snapshots and block storage volumes can be used interchangeably.

Boot configuration

Cloud GPU VMs support flexible boot device configuration, allowing you to modify boot settings through the Cloud API. Select your preferred boot device from attached storage volumes to match operational requirements.

Note: The platform supports only IONOS Linux images at launch.

GPU specifications

The following table provides the specifications:

Specification	Details
Hardware architecture	Utilizes "PCIe passthrough" architecture to provide direct hardware access for optimal performance, simplified deployment, and accelerated production readiness.
GPU model	Primary Offering: High-End NVIDIA H200 GPUs
Maximum GPUs per template	8x GPU units per server
Deployment density	Optimized for high-performance inference workloads

Data security

IONOS provides SSD premium as the default attached volume for Cloud GPU VMs.