"AMD GPU Support" requirements

[dgdo]

Hello everyone,

I have a request from a customer who wants to use "amd gpu support" in some of his vms for his application, based on imagemagick (https://imagemagick.org/script/opencl.php).

We use only server hardware (e.g. DELL R640) without built-in graphics card(s).
The CPUs on all nodes are: Intel(R) Xeon(R) Gold 5120 CPUs.

• What exactly do we need to provide the customer with the desired "AMD GPU Support" in their VMs on the cluster?
  • Does the (exist any?) Server CPU have to include a GPU?
  • as far as i know, this is not the case with server CPUs...
• Do I have to install an AMD graphics card (a specific one?) for this?

As far as I know, you can assign a corresponding device to a VM via PCI passthrough.

• Can the corresponding "pci-device" then be assigned to several VMs or is it then dedicated to just one VM?
• Do you know a server that meets the requirements?

It would be great if you could support me in giving the customer an appropriate environment.

Thanks
Mehmet

 

[leesteken]

Proxmox does have PCIe passthrough for devices to one VM, but this can be hit or miss and depends very much on the IOMMU groups of the motherboards and BIOS.
Sharing PCIe devices over multiple VMs in the way above requires SR-IOV but not many (GPU) devices support that (and I have no experience with this).
Sharing GPU capabilities over multiple VMs is called mediated devices and I do believe Proxmox support GVT-g, but I don't think AMD has support for either of that.

Sorry for not really answering your question, but hopefully my answer gives you some pointers to things to look up and consider.