We are exploring the use of NVIDIA GPUs, but from the context of making the GPU resources available within Kubernetes.   Currently we're trying to get the NVIDIA GPU Operator for Kubernetes working, and we haven't done much testing of PCI passthrough to VMs.

It's interesting that the NVIDIA documentation you linked doesn't seem to mention the "Hopper" architecture at all.

Can you provide a source for the nvidia-vgpu-ubuntu-aie-580_580.82.0.deb package that you said you installed?

After installing the NVIDIA vGPU software were the NVIDIA-specific vfio drivers installed as per Virtual GPU Software User Guide - NVIDIA Docs ?

# lsmod | grep vfio
nvidia_vgpu_vfio       27099  0
nvidia              12316924  1 nvidia_vgpu_vfio
vfio_mdev              12841  0
mdev                   20414  2 vfio_mdev,nvidia_vgpu_vfio
vfio_iommu_type1       22342  0
vfio                   32331  3 vfio_mdev,nvidia_vgpu_vfio,vfio_iommu_type1
#


Which version of the guest drivers were you using?

Regards,
Chris

On 10/30/2025 6:07 AM, GIUSEPPE DEL GAUDIO wrote:
Dear StarlingX community,

We are attempting to configure VGPU sharing via MIG (Multi-Instance GPU) on NVIDIA H200 cards within a StarlingX (STX) 24.09 environment running OpenStack (Caracal).

Platform's configuration details:
      StarlingX (STX) 24.09 con OpenStack(Caracal).
      Hardware: GPU NVIDIA H200.
      CPU : AMD EPYC
      Kernel : vmlinuz-6.6.0-1-amd64

We successfully installed and deployed the nvidia-vgpu-ubuntu-aie-580_580.82.0.deb patch on the two controllers via USM.
The NVIDIA platform is functional: we were able to perform GPU segmentation using MIGs and associate these cuts with the VFs, following the official NVIDIA GRID VGPU guide (https://docs.nvidia.com/vgpu/latest/grid-vgpu-user-guide/index.html#ubuntu-install-configure-vgpu). The GPU supports VGPU/MIG sharing via VFs (Virtual Functions) using Vendor-Specific VFIO.

To enable VF sharing with OpenStack, we applied the following helm override configuration in StarlingX, updating Nova parameters based on the official OpenStack PCI Passthrough documentation (https://docs.openstack.org/nova/latest/admin/pci-passthrough.html)


where 0000:03:00.2 it is the PCI address of a VF linked to a MIG with capacity of 7G_141.

However, when we attempt to attach the VF to a Guest VM using a Nova flavour, the physical node crashes, resulting in the attached logs.

Does anyone have specific experience or documentation regarding this configuration of StarlingX/OpenStack, VGPU/MIG with NVIDIA Vendor-Specific VFIO} ?


Specifically, we are looking for the following:

Note: 
    A VM without GPU but with ethernet card shared via SRIOV works well

Thank you ! 

Giuseppe Del Gaudio - Cloud Engineer | Cloud & Digital Architecture

Descrizione: Descrizione: Descrizione: Descrizione: cid:image001.gif@01CD1BCB.F08213A0

 

Via Bastioni 14/ S.Michele 10, Salerno, Italia

Email: giuseppe.delgaudio@nttdata.com

Tel: +39 3346368189

Learn more at www.nttdata.com/it




NTT DATA Italia S.p.A. - Società per azioni, soggetta ad attività di direzione e coordinamento di NTT DATA EMEA Ltd. - Sede legale a Milano, in Via Ernesto Calindri, 4 – codice fiscale e iscrizione al Registro delle imprese di Milano Monza Brianza Lodi n. 00513990010, iscritta al Repertorio economico Amministrativo n. 974124 - Capitale sociale di euro 40.970.700,00, interamente versato.