VMware vSphere

 View Only

Nvidia vGPU GRID profiles - Resource allocation question

  • 1.  Nvidia vGPU GRID profiles - Resource allocation question

    Posted Oct 19, 2024 09:38 AM
    Edited by a_p_ Oct 20, 2024 07:36 AM

    Dear all,

    I can start less VMs with vGPU GRID profiles than supposed to.

    The server config looks like this:
    1x HPE DL380 Gen11 servers with 2x Nvidia L40s cards
    Each card has 48GB of vGPU memory, so 96GB in total

    Graphics Device Setting for each Nvidia Card: "Shared Direct"
    Host Graphics Setting: "Shared direct" + "Spread VMs across GPUs"

    Nvidia VIB on the host: 535.183.04-1OEM.702.0.0.17630552
    vCenter + ESXi version: 7 Update 3o


    VM config:
    Additional PCI device -> GRID profile: nvidia_l40s-16q or nvidia_l40s-8q

    I configued 3 VMs with each the nvidia_l40s-16q profile.
    Those are up and running. So 48GB are consumed, 48GB are still available.

    When I try to start additional VMs with the nvidia_l40s-8q profile, I can only start one of them. The others fail with the following error:

    "No host is compatible with the virtual machine" "Insufficient resources. One or more devices (pciPassthru0) required by VM vm5 are not available on host host01."

    In my opinion I should be able to start additional 5 VMs with the 8q profile, right?!

    Besides the old Nvidia GRID vib, did I configure something wrong?

    Just browsed the release notes of version 8U3 - are those mixed profiles per card like I want (8q +16q) not possible before 8U3??

    Thanks for your input!