Compare NVIDIA vGPU, Native MIG, Hyper-V GPU sharing and GPU passthrough for CAD workstations, AI inference, Windows VMs and multi-GPU workloads. Review resource sharing, licensing and cost factors to find a setup that fits your applications, operating system and workload size.
Need to centralize GPU desktops, keep AI jobs from competing for resources, share a GPU across Windows VMs, or give one VM full-GPU access? Compare the options below.
Move CAD, Revit, and engineering desktops to a GPU server and deliver them remotely. Each VM gets a defined GPU memory profile, so teams can share a server instead of buying a separate workstation for every user.
Example scenarios:Revit model reviews, CAD part design, remote design teams and hands-on software training.
Explore related products →Run separate AI inference or development jobs on one supported GPU. MIG assigns each job a hardware-backed compute and memory slice, helping prevent one job from using another’s allocated resources. Linux only.
Example scenarios: Product image sorting, call transcription and separate AI model tests.
Explore related products →Use Easy-GPU-PV/WDDM to share GPU resources across Windows VMs for training labs, software testing, or video workloads. GPU resources are shared dynamically, so a busy VM can affect others.
Example scenarios: Windows software training, GPU app testing and shared video processing.
Explore related products →Pass one or more physical GPUs through to a VM for model training, rendering, simulation, or AI inference GPU hosting. The VM gets exclusive use of its assigned GPU while keeping its own operating system and driver environment.
Example scenarios: AI model training, 3D animation rendering, engineering simulations and multi-GPU research.
Explore related products →Host OS runs on the physical server. Guest OS runs inside a VM. Native MIG can also run workloads directly, without a VM.
| Key Difference | NVIDIA vGPU | Native MIG | Hyper-V GPU-PV(WDDM) | GPU Passthrough |
|---|---|---|---|---|
| Resource Allocation | Each VM gets a set amount of GPU memory. GPU compute time is shared under the chosen scheduling policy. | Each instance gets its own compute and GPU memory resources. | Windows VMs share GPU compute and memory through WDDM. Each VM does not have its own dedicated hardware slice. | One VM has full use of one or more GPUs. |
| Main Limitation | Requires a paid vGPU license and a supported platform. | Each job must fit its selected MIG profile. Applications that need multiple GPUs working together require a separate compatibility check. | No hardware-level fault isolation; one VM's GPU fault may affect neighbours. GPU resources are shared rather than reserved for each VM. | The assigned GPU cannot be shared with other VMs. Live migration is generally unavailable with standard GPU passthrough. |
| Host OS / Platform | Supported Linux KVM / Proxmox VE / VMware ESXi setup or Windows Server 2025 + Hyper-V. Contact us to confirm platform compatibility. | Supported Linux distribution. | Windows with Hyper-V. | Compatible Linux, VMware ESXi or Windows Hyper-V platform with GPU passthrough support. |
| Guest / Workload OS | Supported Windows or Linux guest. | Linux guest or bare-metal Linux | Compatible Windows guest. | Supported Windows or Linux guest. |
| Recommended Product | GPU Dedicated Server | GPU VPS / GPU Dedicated Server | GPU Dedicated Server | GPU Dedicated Server |
OS compatibility depends on the GPU, driver and virtualization platform. GPU-PV here refers to Easy-GPU-PV / WDDM sharing. It is different from the SR-IOV-based GPU-P feature in Windows Server. Contact us to confirm your configuration.
Tell us what you need to run. Find a GPU setup to consider, see the main costs and compare your options.
Use this result as a starting point. We will check the GPU, software and capacity for your applications before quoting a setup.
NVIDIA licensing: NVIDIA Virtual GPU (vGPU) software requires a paid license. Native MIG, standard GPU passthrough, and the Easy-GPU-PV / WDDM setup described here do not require an additional NVIDIA vGPU license; other licensing terms still apply. NVIDIA licensing details.
Microsoft licensing: Windows licensing for the host, VMs and remote access depends on your setup and the licenses you already have. Microsoft licensing guidance.
Choose a GPU for virtualization, then check the hardware and setup with our team before ordering.
Linux applications. Check the MIG profile and assigned resources for your VPS or dedicated server.
Multi-GPU memory is shown per card. It is not one combined memory pool.
Compare all available configurations and pricing?
View More GPU PlansCustomer deployments and workload tests, with key configurations and original evidence.
RTX A6000 48GB / Proxmox VE / Windows guests / 8GB vGPU profiles
An architecture team needed six GPU-enabled design workstations within a 20-VM environment. We resolved the host configuration issue and assigned GPU profiles only to the design VMs, leaving lightweight tasks without a vGPU.
H100 PCIe 80GB / Ubuntu 24.04 / Docker + PyTorch / 1g.10gb
Run separate AI inference services and Linux compute jobs on one supported GPU. Each MIG instance has its own compute and GPU memory resources.
NVIDIA A40 48GB / Windows host / Hyper-V / 8 vCPU + 12GB RAM per VM
A Windows gaming deployment needed multiple VMs without a separate GPU for each one. WDDM GPU sharing, reduced graphics settings and a 24 FPS cap helped control combined rendering demand.
Linux host / KVM + QEMU / IOMMU + VFIO / Windows or Linux guests
A GPU VPS platform needed separate environments for tenants running different operating systems and drivers. IOMMU and VFIO assign physical GPUs directly to individual KVM VMs, with each tenant using its own guest OS and dedicated GPU allocation.
Match your workload to a method, confirm your GPU, and we’ll help you deploy.
List your applications, OS and how many users or tasks will use the GPU at once. Then compare vGPU, MIG, GPU-PV(WDDM) and passthrough by resource sharing and licensing.
Choose a GPU model and count from the available plans, and review the CPU, RAM and storage. Not sure what fits your workload? Contact us and we’ll help you choose.
Confirm the GPU, virtualization method and plan. If you’d like to test it first, you can request a trial by following the process in our FAQ.
Your GPU server is deployed automatically after payment. If your virtualization setup needs manual configuration, our technical team is available 24/7 to assist.
Looking for more GPU options? Explore all GPU plans.
If you need to create and manage more than one GPU virtual machine, choose a GPU Dedicated Server. It lets you choose the VM software, create VMs and assign GPU resources. A GPU VPS is a ready-to-use VM, so it is a simpler choice if you just want to install and run applications.
Proxmox is not offered as a GPU VPS OS option. For this setup, choose a GPU Dedicated Server. If all 10 VMs need GPU access, evaluate supported NVIDIA vGPU hardware rather than assuming an RTX 5090 can be split between them. VM capacity and Windows licensing need to be confirmed before ordering.
Nested virtualization is available only on GPU VPS configurations running on our Ubuntu 24 host platform. Installing Ubuntu inside a guest does not enable this capability by itself. Additional VMs share the VPS’s existing resources and add overhead; for your own multi-VM environment, a GPU Dedicated Server is the recommended starting point. Contact us to confirm compatibility.
When choosing a GPU for CAD, consider NVIDIA vGPU if several users need virtual workstations for Revit, drafting or 3D design. Each VM gets a set amount of GPU memory while sharing compute resources. Choose supported hardware and an appropriate NVIDIA RTX Virtual Workstation license, then choose a profile based on model size, detail and how many people work at once.
Yes. NVIDIA vGPU can provide virtual workstations for CAD, design and engineering software training. Hyper-V GPU-PV is another option for supported Windows labs where VMs can share resources as needed. Plan for the students working at the same time. The GPU size also depends on the software and the tasks in each lesson.
Both let multiple Hyper-V VMs share a GPU. Microsoft’s SR-IOV-based GPU-P requires supported hardware, drivers and licensing. The Easy-GPU-PV / WDDM setup described on this page shares GPU resources without dedicated hardware slices, so heavy use by one VM can affect others. Contact us to find a setup that fits your applications.
Yes. With MIG on supported GPUs (such as A100 or H100), one GPU is split into isolated instances, each with dedicated compute and memory, which suits production workloads that need stable performance. For lightweight workloads or GPUs without MIG, multiple services can share the GPU through time-slicing: simpler and more efficient, but heavy loads may interfere with each other.
Consider a GPU Dedicated Server with whole-GPU passthrough when one VM needs exclusive access to multiple GPUs. The setup must support your AI software, model size and the way your GPUs share data. Adding the memory capacities of several cards does not automatically create one memory pool; the application must support distributing the workload.
Yes, we provide a 24-hour free trial for new clients. Here’s how to request one:
1.Choose a plan and click 'Order Now'.
2.In the notes section, type '24-hour free trial' and click 'Check Out'.
3.Click 'Submit Trial Request' at the top right, and complete your personal information. No payment is required.
Share your applications, VM requirements and preferred GPU. Our team will help you review the deployment options and find a suitable configuration.
Have a plan in mind? Ask about a trial when you contact us.
Ready to choose your own setup?
Browse GPU Plans →
