Hardware virtualization requirements
Each workspace is a OpenShift Virtualization virtual machine. OpenShift Virtualization uses the hardware virtualization of the node (Intel VT-x or AMD-V through KVM), so the worker nodes that run the workspace VMs must be bare metal:
On premises: OpenShift Container Platform installed on bare-metal servers, with virtualization enabled in the firmware.
Public cloud: bare-metal instance types, for example
m5.metal,m6i.metal, orc5n.metalon Amazon Web Services (AWS). Virtualized instance types cannot run the VMs, because nested virtualization is not supported.
The control plane nodes can be virtual machines or virtualized instances. For the full list of requirements, see Preparing your cluster for OpenShift Virtualization.
Sizing requirements
Size the bare-metal workers for the shared services plus one VM per user. Each VM requests its full memory from the node, so the amount of memory on the node limits the number of workspaces the node can run.
| Component | vCPU | Memory | Storage |
|---|---|---|---|
Shared services: Operators, Argo CD, Keycloak and its database, Vault, governance interceptor | 8 | 16 GiB | 50 GiB, including the template VM image |
Each workspace VM (default, set in the | 4 | 8 GiB | 40 GiB |
| Node type | Number of nodes | Example (AWS) |
|---|---|---|
Control plane | 3 |
|
Bare-metal worker | 2 |
|
Two bare-metal workers of this size hold the shared services and dozens of workspaces. With smaller servers, plan for the shared services plus 4 vCPU and 8 GiB of memory per user. Ensure that there is room to move VMs from one node to another in case of node failure.
Storage requirements
The cluster needs a default storage class for the VM disks (ReadWriteOnce, 40 GiB per user by default) and for the template image that the disks are cloned from. Storage that supports smart cloning, such as Red Hat OpenShift Data Foundation or a CSI driver with volume snapshots, makes new workspaces start faster.
Model serving
By default, the agents use NVIDIA Nemotron on build.nvidia.com and need no GPU in the cluster. If you serve models on the cluster instead, for example with vLLM, add the GPU nodes that the model server needs. For more information, see Ideas for customization.
