As I’ve mentioned before, the NCasT4_v3 is probably the sweet-spot GPU enabled Azure desktop. It offers excellent performance for a reasonable price, and can handle most mid level CAD and modelling duties.
I’ve no idea of the reasons, but there seem to be a lot of caveats specifically for it, which can make taking advantage of it a challenge. Specifically, you need to manually install the GRID drivers to enable 3D acceleration…
“The Azure Nvidia GPU driver extension will deploy CUDA drivers on the NCasT4_v3-series VMs. For graphics and visualization workloads manually install the GRID drivers supported by Azure.”
“The GRID drivers redistributed by Azure do not work on non-NV series VMs like NCv2, NCv3, ND, and NDv2-series VMs. The one exception is the NCas_T4_V3 VM series where the GRID drivers will enable the graphics functionalities similar to NV-series.”
This took me a while to discover, as I prefer to automate the driver install process using extensions. I spent a couple of hours checking policy settings before I realised this!
Great, so we download the current supported GRID drivers and we are good to go? Well, sadly not in my case.
Running the driver installer resulted in the below message. Annoyingly, the installer deletes its temp files as soon as you close this window.
So before clicking Cancel, navigate to c:\NVIDIA and copy the DisplayDriver folder to your desktop.
Next, open Device Manager and Uninstall the listed Nvidia card. Select ‘attempt to delete drivers’ as well.
Rescan the hardware, and you will see an unknown Display Controller. Right-click this and choose Update Driver, then browse to the DisplayDriver folder you copied to your desktop and click OK.
Strange things happen…
…but the result is, you will finally see the GPU that you are paying for!
What’s more, you will now be able to run the GRID driver install routine successfully.
Once complete, reboot your machine and you should have full GPU acceleration available.
I realise this blog is a super-edge case, and not many organisations are taking advantage of GPU based Azure desktops – but when configured correctly they work really well, so I hope this may be of use to someone.
Thanks for reading!
Credit to David Li for his Fluid Particles (david.li) demo, a very useful tool for quickly assessing graphical performance on a VM without needing to install any software.