Overview
The On-Prem AI Infrastructure Platform is vendor-neutral. NVIDIA GPUs, Intel Gaudi HPUs, and AMD GPUs run side by side on the same platform. Each vendor gets its own compute flavor, Kubernetes cluster template, and accelerator operator, so every workload runs on the accelerator that fits it best.Accelerators are presented through PCI passthrough at the compute layer,
and a vendor-specific operator runs on each Kubernetes node group.
Vendor Model
Architecture
Vendors
- NVIDIA
- Intel Gaudi
- AMD
NVIDIA GPUs are available as dedicated passthrough devices or as shared vGPUs. In
Kubernetes, the NVIDIA GPU Operator manages drivers and runtime, MIG partitioning,
time-slicing, and DCGM monitoring.
Benefits
No vendor lock-in
Add accelerators from any supported vendor without changing how workloads are deployed.
Right accelerator per workload
Match training, inference, and development workloads to the accelerator that suits them.
Clean separation
Separate flavors and templates keep each vendor’s drivers and operators isolated to its
own node groups.
One operating model
Every vendor is managed through the same console, flavors, and cluster templates.
Next Steps
GPU and HPU Performance Optimization
Tuning and telemetry for each vendor’s accelerators.
Unified Console
Manage every vendor’s workloads from one portal.
GPU-Capable VMs and Containers
Passthrough, vGPU, and SR-IOV in detail.
Cluster Templates
How cluster templates are defined in Polystack K8SaaS.