The following figure shows the architecture of the validated solution, focusing on the compute layer and how the graphics components are divided between the VDI and AI workloads.
This architecture aligns with the VMware Horizon block/pod design in which a pod is divided into multiple blocks (as outlined in the following documents: Design Guide and Validation Guide). Each block is made up of one or more vSphere clusters and an associated vCenter server appliance. The figure demonstrates how a single compute node is logically divided to support up to 24 users and a single AI virtual machine (VM) per VxRail.
The following figure shows a design where VDI and compute workloads run in parallel. When these workloads run consecutively, each workload has 3 GPUs assigned to it when running.
The deployment option for this solution supports all of the cloning techniques available from VMware: full, linked, and instant.
A VMware vSAN-enabled vSphere Cluster can have a maximum of 64 nodes and 6,400 VMs per cluster. To expand from this limit, you can add clusters and balance the VMs and nodes across the new clusters.