NVIDIA-Certified Hypervisors#
Introduction#
AI infrastructure is increasingly delivered on virtualized platforms. Hypervisor providers and infrastructure independent software vendors (ISVs) are enabling GPU-enabled virtual machine environments for compute and AI workloads, including training, inference, and emerging agentic AI use cases. Enterprise customers and cloud providers look for validated platforms that preserve virtualization benefits without compromising performance.
NVIDIA-Certified Hypervisors is a certification program for GPU-optimized, enterprise-grade virtualization infrastructure platforms that provide a hypervisor layer for GPU-backed virtual machines (VMs), extending the NVIDIA-Certified portfolio to the virtualization layer. The program uses NVIDIA hardware, validation suites, and control-stack baselines to verify that tested configurations expose NVIDIA GPUs correctly to guest VMs and deliver near bare-metal performance for virtualized compute and AI workloads.
Designed to evolve with NVIDIA GPU architectures and GPU server platforms, the program provides NVIDIA Cloud Partners and enterprise customers with a clear, trusted signal for identifying hypervisor platforms that have been tested against NVIDIA-defined requirements as their AI infrastructure evolves.
Certified configurations are available in the List of NVIDIA-Certified Hypervisors, with the applicable partner platform, NVIDIA system, GPU architecture, and guest VM configuration. Customers can use these listings to identify validated partner offerings for their deployment requirements. Partners can use them to communicate the specific configuration covered by certification.
Virtualization with NVIDIA GPUs for AI and Compute#
Virtualization remains important for accelerated infrastructure because it provides operational capabilities that bare-metal deployments do not always address on their own. These include multi-tenancy, workload isolation, standardized provisioning, policy enforcement, lifecycle management, and Day 2 operations across shared infrastructure. These capabilities help enterprises and cloud providers operate GPU-enabled environments with greater consistency as deployments move from pilots to production AI factories.
For NVIDIA GPU-backed virtual machines, deployment models can vary based on workload, tenancy, and operational requirements.
Deployment Model |
Details and Program Scope |
License Requirements |
|---|---|---|
GPU passthrough |
Assigns one or more physical GPUs, along with associated device paths, to a guest VM. Current certification scope for listed configurations. |
No NVIDIA AI Enterprise license required. Uses the NVIDIA Data Center Driver in the guest VM. |
Enables fractional, 1:1, and multi-GPU VM configurations. Potential future certification coverage; not part of current scope. |
Requires an NVIDIA AI Enterprise license. |
Note
License requirements refer to NVIDIA software licensing only; partner platform or hypervisor licensing is handled separately.
Program Overview#
NVIDIA-Certified Hypervisors is a performance certification program for partners whose hypervisor platforms or virtualization infrastructure software expose NVIDIA GPUs to guest VMs for compute and AI workloads.
For this program, a certifiable platform is a partner-defined and supported virtualization software stack that creates, runs, and manages virtual machines with NVIDIA GPUs and associated devices exposed to the guest VM. A certified configuration includes the hypervisor, host OS or kernel, device-assignment model, management layer, and NVIDIA software baseline.
The program validates platform behaviors that are most critical to virtualized NVIDIA GPU performance, as outlined in the Optimizing VM Configuration for Performant AI Inference whitepaper.
The program validates:
GPU assignment: Confirms that NVIDIA GPUs and associated PCIe devices are exposed correctly in the guest VM.
Topology visibility: Confirms that the guest VM can see the expected GPU, network interface card (NIC), NVLink, NVSwitch, PCIe, and non-uniform memory access (NUMA) relationships.
Data-path locality: Confirms that GPU-to-NIC locality and GPUDirect RDMA paths are preserved where applicable.
NVIDIA software baseline: Confirms that the partner stack delivers expected performance against NVIDIA’s virtualized control-stack baseline.
Certification applies to the tested partner stack version, NVIDIA GPU architecture, hypervisor configuration, guest VM configuration, validation workload scope, and NVIDIA software baseline.
NVIDIA intends to expand NVIDIA-Certified Hypervisors certification coverage over time to additional deployment scenarios and platforms based on future NVIDIA CPU and GPU architectures, including NVIDIA Vera and NVIDIA Rubin. These scenarios are outside the current certification requirements until NVIDIA publishes the associated validation scope and supported configurations.
Key Objectives#
The NVIDIA-Certified Hypervisors program is designed to:
Validate performance for virtualized NVIDIA GPU workloads against NVIDIA-defined requirements.
Enable virtualized infrastructure platform vendors to validate their GPU-enabled virtualization platforms and infrastructure stacks on the latest NVIDIA GPU architectures.
Establish a repeatable validation and evidence framework for virtualized AI and accelerated computing workloads, including model-serving workloads used by generative and emerging agentic AI applications.
Provide enterprise customers and NVIDIA Cloud Partners with a clear, trusted signal to identify validated, GPU-optimized hypervisor-based infrastructure platforms.
Extend certification coverage across current and future NVIDIA x86 and Arm-based systems, NVIDIA GPU architectures, and GPU server platforms.
Certification Tracks#
NVIDIA designates two certification tracks so the program can accommodate hypervisor partners with different platform strategies and deployment targets across current and future NVIDIA platform architectures.
Arm Certification#
The Arm certification track is designed for hypervisor platforms that virtualize NVIDIA Arm-based accelerated systems for running AI and compute workloads.
The current certification scope validates a single NVIDIA GB200 NVL compute-tray configuration. Over time, NVIDIA may expand the track to additional NVIDIA Arm-based system architectures and configurations as associated validation requirements and supported configurations are published.
Scope |
Configuration |
What It Demonstrates |
|---|---|---|
Single Grace Blackwell Compute Tray |
One 4-GPU / 2-CPU passthrough VM spanning one GB200 NVL compute tray. |
The partner stack can expose Grace CPU and Blackwell GPU resources to the guest VM and preserve compute-tray performance. |
x86 Certification#
The x86 certification track is designed for hypervisor platforms on x86-based NVIDIA accelerated systems that support full-GPU passthrough virtual machines for AI and compute workloads. The track is intended to evolve across Hopper, Blackwell, and future NVIDIA GPU architectures.
x86 Certification Phases#
The x86 certification track is organized into two required phases that together define the current x86 certification scope. Partners are expected to complete both phases to be certified as an NVIDIA-Certified Hypervisor on the x86 track.
Phase |
Configuration |
What It Demonstrates |
|---|---|---|
Single-Node |
One 8-GPU passthrough VM on a single x86 Hopper SXM-based system. |
The partner stack can expose a full 8-GPU node to the guest VM and preserve single-node GPU performance. |
Multi-Node |
Two 8-GPU passthrough VMs deployed across two x86 Hopper SXM-based systems. |
The partner stack can preserve multi-node GPU communication and data paths for distributed compute and AI workloads. |
Both certification tracks evaluate whether the partner’s virtualized stack can:
Expose NVIDIA GPUs and associated PCIe devices correctly to the guest VM.
Preserve guest-visible GPU, NIC, NVLink, NVSwitch, PCIe, and NUMA topology.
Maintain GPU-to-NIC locality for high-performance data paths such as GPUDirect RDMA.
Deliver near bare-metal performance against NVIDIA control-stack baselines.
Partners may support additional VM sizes, PCIe GPU servers, or other deployment models as part of their commercial product offerings. NVIDIA-Certified Hypervisors certification applies only to the configurations validated and listed by NVIDIA. Over time, NVIDIA may expand certification tracks to additional supported VM configurations and deployment models.
Validation Suite#
NVIDIA-Certified Hypervisors uses an NVIDIA-defined validation suite to evaluate the platform behaviors that are most critical to virtualized compute and AI workloads. The suite combines topology evidence, low-level platform benchmarks, GPU fabric tests, and representative AI workload performance.
Category |
Focus |
Representative Metrics |
Use Case |
|---|---|---|---|
GPU Collective Communication |
GPU-to-GPU communication within a VM and, where applicable, across nodes. |
Communication bandwidth, throughput. |
Distributed AI training and fine-tuning, multi-GPU inference, and communication-intensive HPC. |
GPU, CPU, and Memory Bandwidth |
Data movement across GPU, CPU, and memory paths in the virtualized platform. |
Bandwidth, streaming throughput. |
Data-intensive AI training and fine-tuning; high-throughput or batch inference, preprocessing, and analytics. |
Compute Performance |
Core GPU compute behavior for AI and HPC-style operations. |
Compute throughput. |
AI training and fine-tuning, HPC simulation and scientific computing, and GPU-accelerated analytics. |
AI Inference |
Real workload behavior for model serving. |
Throughput, latency, and time-to-first-token. |
Gen AI, high-throughput and batch model serving, large-scale large language model (LLM) and multimodal inference, and agentic AI applications. |
RDMA Data Path |
Cross-node GPU communication over the network fabric where applicable. |
GPU-to-network throughput and RDMA data-path performance. |
Multi-node AI training, distributed large-scale inference, MPI, and network-bound HPC. |
Topology and Device-Mapping Evidence |
Guest-visible device relationships that support performant GPU workloads. Supports validation review, but is not treated as a workload category. |
Confirms guest-visible GPU, NIC, PCIe, NVLink, NVSwitch, and NUMA relationships. |
Cross-cutting evidence for multi-GPU/multi-node training, scale-out inference, and HPC; not a workload category. |
Note
Use-case mappings indicate the workload patterns for which each test category provides relevant evidence. No single test category represents end-to-end application performance.
The validation suite is tied to the certification track, phase, NVIDIA GPU architecture, and guest VM configuration being tested.
Ongoing Expectations#
NVIDIA-Certified Hypervisors certification applies only to the configuration validated by NVIDIA, which includes the following:
NVIDIA GPU architecture.
Partner platform/software stack version.
NVIDIA driver, firmware, and control-stack baseline.
Guest VM configuration, including number of GPUs/CPUs, memory sizing, device assignment model, and other attributes.
Validation suite version.
Topology requirements for the certified configuration.
Within the same NVIDIA GPU architecture, closely related GPU SKUs may be treated as equivalent for certification when hardware, software, topology, and virtualization-relevant behavior are not materially different. NVIDIA standardizes certification on a reference GPU SKU and validated platform configuration for each architecture to keep validation consistent, repeatable, and aligned to the representative hardware capabilities of that generation.
Certification remains valid for the validated configuration unless the partner makes a material change to the certified hypervisor stack or configuration. Material changes may include a major partner hypervisor stack release, changes to the guest VM configuration, changes to the device assignment model, or changes that affect guest-visible topology or required data paths.
A new NVIDIA GPU architecture, platform, or certification scope is treated as a new certification, not a recertification of an existing configuration, because platform capabilities, software baselines, memory capacity, and validation recipes may change.
Partners interested in NVIDIA-Certified Hypervisors should review this document and engage NVIDIA to begin onboarding to the program. NVIDIA will work with partners during onboarding to review certification fit, target platform scope, and the appropriate certification track.