Our primary workload is LLM inference, and we want to understand the available GPU virtualization capabilities in XCP-ng before proceeding with the deployment.

We would appreciate guidance on the following:

1. NVIDIA H200 Compatibility

• Does XCP-ng 8.3 support NVIDIA H200 GPUs?
• Are there any known compatibility issues with H200 PCIe passthrough?
• Is there any specific configuration required for IOMMU, PCIe passthrough, BIOS or GPU drivers?

2. GPU Passthrough

• Can we dedicate an entire NVIDIA H200 GPU (141 GB HBM3e VRAM) to a guest VM?
• Are there any recommended configurations or installation guides for NVIDIA data center GPUs?

3. NVIDIA vGPU and VRAM Allocation

• Does XCP-ng support NVIDIA vGPU technology with H200 GPUs?
• Can we partition a physical H200 GPU and allocate specific VRAM capacities to multiple guest VMs?

For example, can we allocate 20 GB, 40 GB or 80 GB of GPU VRAM to individual VMs?

Does XCP-ng support NVIDIA MIG (Multi-Instance GPU) on H200, or is there any alternative approach to achieve GPU sharing?

4. GPUaaS Architecture
   Our objective is to build a multi-tenant GPUaaS platform where customers can provision VMs with dedicated or partitioned GPU resources for LLM inference.

We would like to understand whether this architecture is achievable with XCP-ng and what the recommended approach would be.

If anyone has successfully deployed NVIDIA H100/H200 GPUs with XCP-ng, we would particularly appreciate your experience, configuration details and recommendations.

Thank you for your guidance!