Post

OpenShift Virtualization 4.22 on NVIDIA HGX

OpenShift Virtualization 4.22 is an NVIDIA-Certified Hypervisor for Hopper HGX. What 8-GPU passthrough proves, and when that GPU stays on a VM versus a pod.

OpenShift Virtualization 4.22 on NVIDIA HGX

Personal site note: views expressed here are my own and do not necessarily represent Red Hat or any other organization. Certification scope moves. Confirm the NVIDIA-Certified Hypervisors list and the Red Hat announcement for the stack you will actually run.

On 21 October 2026, Red Hat OpenShift Virtualization 4.22 joined the NVIDIA-Certified Hypervisors list. The row is specific: Red Hat OpenShift Container Platform 4.22 with OpenShift Virtualization 4.22, on the x86 track, for NVSwitch-based 8-GPU NVIDIA Hopper SXM systems. Red Hat’s announcement names that board class as NVIDIA HGX.

OpenShift Virtualization is on that list as a Kubernetes-native hypervisor: the virtual machine is a KubeVirt workload on KVM, scheduled like any other pod, and the GPUs are assigned by passthrough.

GPU virtual machines and containerized AI share that cluster as separate GPU node pools. The listing is evidence about an 8-GPU passthrough guest, its topology, and representative inference. It is the operations piece that sits beside What’s New in Red Hat OpenShift AI 3.5 and Confidential AI on OpenShift: TEEs, GPUs, and Trustee.

What the listing actually covers

NVIDIA-Certified Hypervisors is a performance certification for a named stack: hypervisor, host OS, device-assignment model, management layer, guest shape, and NVIDIA software baseline.

The x86 track has two required phases. The single-node phase is full GPU and NIC passthrough. The multi-node phase repeats that guest, one per node, and checks the path between them.

PhaseGuest the suite runsWhat it demonstrates
Single-nodeOne 8-GPU VM on one Hopper SXM nodeThe guest sees the node and holds single-node GPU performance
Multi-nodeTwo 8-GPU VMs, one per node, on two identical nodesCross-node GPU communication and the RDMA data path still work

The current certification scope for that track is GPU passthrough. The guest uses the NVIDIA data center driver. NVIDIA’s program notes say this model does not require an NVIDIA AI Enterprise license. NVIDIA vGPU—fractional, 1:1, or several virtual GPUs per physical GPU—is called out as possible later coverage. It is outside the listing that published today, and it does require an NVIDIA AI Enterprise license. OpenShift and other platform subscriptions are a separate contract.

The Arm track is a different table: a 4-GPU, 2-CPU passthrough VM on one Grace Blackwell compute tray. OpenShift is on the Hopper SXM table. A GB200 design is a new certification when NVIDIA lists it, which matches how the program treats a new GPU architecture.

Partners may sell other VM sizes and PCIe GPU servers. Those offerings can be supported. They are certified only when they appear on the list. A material change to the hypervisor stack, the guest shape, or the device-assignment model can require a new review. Treat a future 4.23 or a Blackwell HGX row as a new line, and read it again before you cite this one.

What the suite measures

The useful part of the announcement is the test list. “Near bare-metal” in the Red Hat post and in NVIDIA’s program description means those tests, on that 8-GPU guest, against NVIDIA’s virtualized control-stack baseline. Your model, batch size, and network still need their own measurement.

CategoryWhat the guest must show
TopologyGPU, NIC, PCIe, NVLink, NVSwitch, and NUMA relationships
CollectivesGPU-to-GPU bandwidth inside the VM, and across nodes in the multi-node phase
CPU, GPU, and memoryBandwidth along those paths
ComputeThroughput for the AI and HPC-style operations in the suite
InferenceThroughput, latency, and time to first token on representative serving
RDMAGPU-to-NIC locality and GPUDirect RDMA where the phase includes it

Topology is the one I would put on a slide. An HGX node is eight SXM GPUs behind NVSwitch, with NICs placed for GPUDirect. NCCL and multi-GPU inference care that the guest sees those links, and that a GPU talks to the NIC that sits next to it. A hypervisor can boot a VM with eight GPUs and still hide or reshuffle that map. The suite checks the map, then runs the bandwidth and inference tests that miss their baseline when the map is wrong.

NVIDIA says no single category is end-to-end application performance. Red Hat’s footnote says the same thing in customer language: run your own proof of concept. I would run the model you will actually serve, on the guest size you will actually deploy, and keep the certified 8-GPU shape as the reference point.

Same cluster, separate GPU nodes

OpenShift Virtualization and the NVIDIA GPU Operator can both be installed on one cluster. The operator then has to be told which kind of GPU node it is building. The label is nvidia.com/gpu.workload.config. One node gets one value.

Label valueWhat lands on the nodeWho can use the GPU
containerData center driver, container toolkit, device pluginPods requesting nvidia.com/gpu
vm-passthroughvfio-pci on the GPUs, sandbox device pluginVirtual machines, one or more whole GPUs
vm-vgpuvGPU manager and mediated devicesVirtual machines with virtual GPUs

An unlabeled node follows the operator default, which is container unless you change sandboxWorkloads.defaultWorkload. A node does not mix these. Passthrough binds the GPU to vfio-pci so the guest driver owns it. A container workload needs the data center driver on the host. Those two bindings do not share a GPU. NVIDIA’s OpenShift Virtualization guidance also assumes the virtual-machine GPU nodes are bare metal.

That is the sentence behind “one cluster.” Build two pools. Label the HGX nodes that will host the certified-style guests vm-passthrough. Label the nodes that will run Red Hat OpenShift AI pods container. Schedule with the usual node selectors. The Kubernetes API is shared. The device is not.

vGPU is a third pool, for density when a physical GPU is larger than the workload. It is a supported OpenShift Virtualization pattern. It is not the pattern this certification measured. Keep it off the slide that cites the Hopper listing.

There is a second supported way to attach a GPU to a virtual machine: OpenShift Virtualization’s own PCI passthrough, with the GPU Operator operands kept off that node so the two stacks do not fight. Use one approach per node. The certification does not relieve you of picking one and writing it down.

When the GPU stays on a virtual machine

Put the GPU on a virtual machine when the guest is the product.

  • The image is an appliance, a Windows build, or an ISV stack that installs its own driver and CUDA, and the host must not own that stack.
  • A VMware GPU virtual machine is moving in an early migration wave and the guest has to stay intact. The migration mechanics are the Migration Toolkit for Virtualization problem; the GPU decision is whether that guest still needs passthrough after it lands.
  • The workload is the certified shape: one guest on a Hopper HGX node, all eight GPUs, NIC passthrough, NVLink and NUMA visible inside the guest. That is the single-node phase. Two such guests on two nodes is the multi-node phase.
  • You want a guest operating-system boundary around a full GPU, and you can spend the whole device on that guest.

Carving that HGX node into 1-, 2-, or 4-GPU virtual machines leaves the listed configuration. Full passthrough of a subset of GPUs also changes NVLink behavior on Hopper, because the NVSwitch fabric is a property of the board. NVIDIA documents a shared-NVSwitch model for smaller guests. Treat that as a separate design, with its own proof of concept, and cite the 8-GPU listing only as the reference result.

The guest still needs CPU and memory sized for the GPUs you handed it, and the NICs that the suite required to be passed through if you care about GPUDirect. A virt launcher pod with eight GPUs and an emulated NIC is a different machine than the one NVIDIA ran.

When the GPU belongs in a pod

Put the GPU on a pod when OpenShift is already the scheduler for the work.

  • Model serving, evaluation, and training are container images: OpenShift AI 3.5, llm-d, Kueue, Ray, or a pipeline step. The day-2 controls for that path—flow control, rollouts, queues—are in What’s New in Red Hat OpenShift AI 3.5. They expect a device plugin, which means a container node.
  • Several workloads should share one physical GPU through MIG or time-slicing. That sharing happens under the host driver. A passthrough guest has the whole device until you shut the guest down.
  • The platform team wants one driver lifecycle on the node, via the GPU Operator, and the application team ships a container that consumes nvidia.com/gpu.

Confidential computing is a pod path of a different kind. A CPU trusted execution environment plus an NVIDIA confidential GPU, attested before weights are released, is the design in Confidential AI on OpenShift: TEEs, GPUs, and Trustee. That runtime is sandboxed containers. This hypervisor listing measures performance of a conventional passthrough guest. Attestation stays on the Trustee design. The performance claim stays on the listing. If the requirement is “the host cannot read the weights,” use Trustee. If the requirement is “this 8-GPU guest performs like the bare-metal Hopper node,” use the certified passthrough shape.

The solutions architect takeaway

Use the 21 October 2026 listing as a boundary, then pick a pool.

  1. Quote the row. OpenShift Container Platform 4.22 with OpenShift Virtualization 4.22, x86 track, Hopper SXM, 8-GPU passthrough. HGX is that board class. Grace Blackwell is a different table.
  2. Quote the guest. One virtual machine owns the eight GPUs and the NICs on a node. Two such virtual machines are the multi-node result. Smaller guests and vGPU are outside this result.
  3. Split the nodes. vm-passthrough for those guests, container for OpenShift AI and other pods. One label per node. Bare metal for the virtual-machine GPU pool.
  4. Read the suite as topology plus inference, then measure your model. The evidence is NVLink, NUMA, GPU-to-NIC locality, bandwidth, and representative time to first token. Your serving stack still gets a proof of concept.
  5. Keep confidential AI on its own track. Attestation is Trustee and a confidential GPU. This certification is performance of a passthrough virtual machine.

Pair the decision with What’s New in OpenShift Virtualization 4.22 for the rest of the release, and with a vendor AI bill of materials only after the GPU Operator and OpenShift versions on that bill match the row you intend to cite. Hardware reference architectures go stale the same way certification rows do; see OpenShift Hardware Vendor Reference Architectures.

Want help placing GPU virtual machines and OpenShift AI pods on one 4.22 cluster? Reach out to your Red Hat account team—or confirm the Hopper HGX row and prove one 8-GPU passthrough guest on a non-prod bare-metal node first.

Further reading

This post is licensed under CC BY 4.0 by the author.