1. Problem Statement

Citrix Virtual Apps & Desktops (VDI) requires a consistent and predictable method for configuring GPU resources when creating VMs from a Master VM template.
Today, GPU configuration in OpenShift Virtualization is vendor-specific (NVIDIA vs. AMD vs. Intel), resource-specific (passthrough vs. mediated/vGPU), and expressed through multiple YAML fields, making it extremely challenging for Citrix to support all hardware types.

Citrix needs a single, vendor-neutral GPU abstraction, similar to Kubernetes StorageClass, so their product can provision GPUs without learning each vendor’s details.

2. Why This Is Needed

Citrix must support on-prem, cloud, multi-vendor, and multi-GPU-type deployments.

The current OpenShift Virtualization GPU configuration requires them to parse and override:

• spec.domain.devices.gpus

• pciHostDevices

• mediatedDevices

• resourceName (e.g., nvidia.com/mig-2g.10gb, amd.com/gpu)

• NodeSelectors (vendor-specific)

This is not maintainable across customer environments. Enterprises expect GPU provisioning to work like storage classes: simple and vendor-agnostic.

3. Proposed Solution (High-Level)

Introduce a GPUClass abstraction that allows users (or Citrix) to specify GPU requirements using a single simple field:

gpuClass: "gpu.vdi.medium

4. Functional Requirements

• A new resource type GPUClass, must allow defining GPU profiles decoupled from hardware/vendor specifics.
• OpenShift Virtualization must map GPUClass → actual hardware configuration (passthrough, vGPU/mdev, MIG slice, resourceName, selectors).
• Citrix Master VM GPU config must be preserved or overridden via Machine Profile using GPUClass.
• The GPUClass abstraction must work across NVIDIA/AMD/Intel

5. Acceptance Criteria

• AC1: A user can define one or more GPUClass resources.

• AC2: A VM using gpuClass: X is scheduled and configured correctly across hardware vendors.

• AC3: Citrix Machine Profile can override master VM GPUClass cleanly.

• AC4: System correctly handles fallback when hardware does not match GPUClass.

• AC5: Documentation includes examples for NVIDIA, AMD, Intel, and cloud GPUs.{}

6. Non- goals

• This RFE does NOT require redesigning device plugins.

• This RFE does NOT replace vendor drivers or MIG tooling.

• This RFE does NOT define performance guarantees.

• This RFE does NOT introduce a new scheduler, only an abstraction layer.