Mirantis k0rdent AI Certified through the NVIDIA-Certified Hypervisors Program

As operators scale AI infrastructure, adding a virtualization layer enables flexible GPU allocation and stronger tenant isolation. However, operators need assurance it won’t come at the cost of the reliable, near bare metal performance that enterprise AI factories and NVIDIA Cloud Partners depend on.

As an inaugural ISV partner in the NVIDIA-Certified Hypervisors program, Mirantis has achieved NVIDIA certification for k0rdent AI's virtualization stack, which delivers predictable, bare metal-like performance on NVIDIA GB200 NVL72 with NVIDIA Grace CPUs. This complements our work earlier this year as an inaugural NVIDIA AI Cloud Ready ISV partner, which validated k0rdent AI as an AI Factory solution for reliability, performance, and operational consistency; Mirantis has now added a validated GPU virtualization layer. 

What the certification covers

NVIDIA-Certified Hypervisors enable customers to confidently choose virtualized infrastructure solutions validated for near bare metal performance across AI and accelerated computing workloads in enterprise AI factories.

The NVIDIA-Certified Hypervisors program consists of a defined battery of tests that are run against a partner's virtualized environment and compared with bare-metal and pure-KVM baselines. The workload-based certification testing assesses representative performance-critical behaviors across compute, memory, data-path efficiency, and LLM inference. Mirantis k0rdent AI has successfully passed all tests for the NVIDIA GB200 NVL72 validation track.

Solutions like this typically take operators years to build and prove out. Mirantis k0rdent AI’s hypervisor certification demonstrates a proven, deployment-ready GPU virtualization layer, backed by Mirantis professional services and support, for operators and enterprises.

Engineered for performance and multi-tenant isolation

The k0rdent AI virtualization stack is topology-aware and optimized for performance. When a hypervisor places virtual machines, it must account for how GPUs, CPUs, network interfaces, and local storage are physically wired together. Workloads may otherwise communicate across NUMA boundaries, eroding performance and introducing security risks. The k0rdent AI virtualization stack maps topology at both the node and rack level, so virtual machines can be placed and networked across NVIDIA NVLink (scale-up) and NVIDIA Quantum InfiniBand or Spectrum-X Ethernet (scale-out) interfaces while preserving optimal performance.

For service providers and enterprises alike, the core benefits are the same: Virtualization no longer necessitates a material performance trade-off, and operators can allocate less than a full node per workload as needed. For operators managing large-scale GPU commitments, this kind of flexibility to match GPU allocation to exact workload needs is becoming a baseline requirement.

Additionally, virtualization enables multi-tenancy, with virtual machines serving as the tenancy boundary. Tenant-isolated resources such as networking, GPU, and storage can be bound to the VM, ensuring fault tolerance and data safety. This provides the necessary primitives to create a hyperscaler-like experience for enterprise clients.

Operational speed and simplicity

Virtualization also changes the operational pace of managing GPU infrastructure. While rebooting a physical machine can take 15-20 minutes, rebooting a virtual machine takes just a couple of minutes. While reassigning a physical node from one tenant to another means hours of reprovisioning it from scratch, reassigning a VM takes just a few minutes to destroy it and spin it up again for a different tenant or workload. For operators running large, multi-tenant GPU fleets, this kind of operational speed and efficiency compounds at scale.

The k0rdent AI virtualization stack also provides a single control plane to manage both VMs and containers, sparing teams from maintaining separate stacks. This is especially important for enterprises, which often run a diverse mix of virtualized and containerized workloads (e.g., AI models in containers alongside more traditional applications like databases and message queues in VMs).

Learn more

k0rdent AI VMaaS is available today as a technical preview. It fits directly into k0rdent AI's infrastructure-as-a-service (IaaS) model, which also includes bare-metal-as-a-service (BMaaS), and container orchestration (i.e., Kubernetes).

To learn more about running virtualized GPU workloads on k0rdent AI, contact us to speak with a Mirantis solutions architect.