19.0-19.5 Only: On GPUs configured with GPU-P for maximum 1-GB vGPU density, some VMs fail to start#

Description#

If a GPU is configured with the maximum number of 1-GB vGPUs allowed on the GPU and each vGPU is assigned to a different VM, some of the VMs fail to start. This issue occurs because NVIDIA vGPU software overstates the frame buffer size of 1-GB vGPUs. For vGPUs with more than 1 GB of frame buffer, NVIDIA vGPU software states the correct frame buffer size.

This issue occurs only with GPU-P on Microsoft Windows Server and Microsoft Azure Local.

Status#

Resolved in NVIDIA vGPU software 19.6

The resolution of this issue reports the correct frame buffer size so that the maximum number of VMs with 1-GB vGPUs allowed on the GPU can start.

The corrected frame buffer size and maximum number of VMs can prevent vGPU migration of VMs between a host that is running NVIDIA vGPU software 19.6 and a host that is running an earlier NVIDIA vGPU software 19 release. To allow migration between such hosts, you can restore the earlier overstatement of the frame buffer size.

Note

If you restore the earlier overstatement of the frame buffer size, the original limit on the number of VMs that can start is also restored. On a host that is running NVIDIA vGPU software 19.6, some VMs fail to start if the GPU is configured with the maximum number of 1-GB vGPUs allowed on the GPU.

Add the DisableVgpuFbCorrection DWord (REG_DWORD) registry value to the Windows registry key HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\Class\{4d36e968-e325-11ce-bfc1-08002be10318}\<instance> and set this registry value to 1.

<instance>
The instance subkey of the GPU, for example, 0000.

To reinstate the correct reporting of the frame buffer size, remove the DisableVgpuFbCorrection DWord (REG_DWORD) registry value or set it to 0.
Reload the NVIDIA vGPU software display driver or reboot the hypervisor host.

Ref. ##

5854376