GPU Support in Private Cloud Director

Unlock high-performance computing with Private Cloud Director's GPU acceleration for virtual machines. Efficiently manage specialized workloads like AI/ML training and video rendering.

Private Cloud Director (PCD) now supports GPU acceleration for virtual machines, enabling high performance computing workloads that require specialized hardware acceleration. This support addresses the growing demand for GPU resources among enterprises, particularly during AI/ML training, video rendering, simulation software, and other compute-intensive applications.

Many enterprise customers need GPU acceleration for their critical workloads, but previously had to manage these resources manually or outside PCD . With GPU support, you can now:

• Assign full physical GPUs in passthrough mode for maximum performance, ideal for demanding ML training jobs, high-end graphics workloads, and applications that require bare-metal GPU performance.
• Share GPU resources efficiently using vGPU mode, allowing multiple virtual machines to access portions of the same physical GPU. This maximizes resource utilization and reduces costs for lighter workloads, such as AI inference and development environments.

The GPU capabilities serve different roles across your organization, each with specific needs and workflows.

Cloud Administrators and Infrastructure Teams configure GPU hosts, manage resource allocation, and monitor GPU utilization across the infrastructure. As a person with this role or similar, you can set up GPU-enabled clusters, create resource policies, and ensure optimal GPU resource distribution.

Developers and Data Scientists consume GPU enabled VMs for their high performance workloads. As a person with this role or similar, you can select appropriate GPU configurations based on their specific needs, from lightweight development and testing to intensive training and inference workloads.

Enterprise Teams running specialized applications benefit from dedicated GPU resources. Engineering teams use GPUs for CAD simulations, media teams for video rendering, and research teams for complex computational tasks.

Before setting up GPU support, familiarize yourself with these essential GPU concepts and terminology.

GPU Passthrough: Assigns an entire physical GPU directly to a single VM. This provides the highest performance, but the GPU cannot be shared with other VMs.

GPU Device Types

Platform9 supports the following GPU device types.

Device type	Description	Used by	Supported on
type-PCI	Standard full GPU passthrough	Passthrough	All supported host operating systems
type-pf	SR-IOV physical function passthrough	Passthrough	All supported host operating systems
type-vf	SR-IOV virtual function passthrough	vGPU	Ubuntu 24.04, Rocky Linux 10 — requires the NVIDIA kernel variant vGPU driver

NOTE

type-vf is used by vGPU only. It is not available as a GPU passthrough option. On non-SR-IOV Ubuntu 22.04 and rocky 10.2 hosts, vGPU uses mediated devices (mdev) rather than type-vf.

vGPU (Virtual GPU): Allows a single physical GPU to be shared across multiple VMs. Each VM receives a "slice" of the GPU with predefined resources.

vGPU Profile: A preset configuration that determines the resources (memory, cores, compute features) allocated to each vGPU slice.

GPU Host Aggregate: A logical grouping of hosts that have GPU capabilities, used to schedule GPU VMs to appropriate hosts.

GPU implementation options

Private Cloud Director supports these GPU modes for different workload requirements:

GPU Passthrough

• Best for: High-performance computing, ML training, applications requiring full GPU access
• Resource model: One physical GPU per VM (exclusive access)
• Performance: Maximum performance (bare-metal equivalent)
• Use cases: Deep learning training, high-end graphics workloads, applications requiring full GPU memory

Learn more about Set up GPU Passthrough

Virtual GPU (vGPU)

• Best for: Multi-tenant environments, development, testing, lighter workloads
• Resource model: Multiple VMs share one physical GPU through virtualization
• Performance: Good performance with resource sharing
• Use cases: AI inference, development environments, graphics virtualization, and cost-effective GPU sharing

Learn more about Set up vGPU

Supported GPU Models

Private Cloud Director supports these NVIDIA GPU models:

GPU Model	Passthrough	vGPU	SR-IOV Required for vGPU
NVIDIA L4*			Yes
NVIDIA L40			Yes
NVIDIA L40S*			Yes
A100*			—
A40			Yes
NVIDIA A2			Yes
NVIDIA GeForce RTX 2080			—
NVIDIA RTX 4000 Quadro			—
Nvidia RTX 4000 Ada Gen			Yes
NVIDIA RTX 6000 Quadro			No
NVIDIA RTX 6000 Ada Gen			Yes
NVIDIA RTX 6000 Pro			Yes
Tesla T4			No
NVIDIA H100			—
Nvidia H200 120GB, 144GB, 480GB			—
NVIDIA A10			No

*Private Cloud Director validated

Non-SR-IOV GPU models (for example, Tesla T4, Quadro RTX 6000/8000) do not require SR-IOV to be enabled in the BIOS for vGPU use. These models expose mediated devices (mdev) directly from the physical GPU function. SR-IOV BIOS settings apply only to SR-IOV-capable GPU models.

Additional Resources

• GPU VM Machine Type and OS Compatibility
• Troubleshooting GPU Support
• vGPU Inventory Troubleshooting and Multiple vGPU per VM
• GPU VM Recovery After Host Upgrade
• GPU FAQs

Platform9 OS Beta deployers

If you are installing PCD via the Platform9 OS ISO, GPU configuration is available directly in the installer wizard as Step 3 of 4. The wizard presents the same Passthrough and vGPU options described above. Refer to Platform9 OS Beta Installation for the full ISO-based installation walkthrough, including the GPU configuration step.