GPU virtualization design help required

We want to deploy VMware Horizon with A100. We want to run AI and ML with parallel computing. Two nos of server each casting two A100 40GB GPU. Our requirement is as below:

  1. Assign full GPU (hosted at two servers) at a single VM
  2. Sharing GPU with multi VM.

What would be the licensing and additional kit requirement to be count for the OEM order and for the Nvidia?

Hey Anis,

For VMs on A100 you will need to use our vCS license. vCS is licensed per GPU so if you have two A100s you will need 2 vCS licenses.

Prereqs for vGPU can be found here: Virtual GPU Software User Guide :: NVIDIA Virtual GPU Software Documentation.

best regards
D

Thanks for the information. Just to make little more understanding, if we set 2 nos A100 per server and we configure 3 host with the same hardware, can we assign all 6 GPU at a single VM managed by VMware Horizon with the vCS license or any additional license or kit may require? As these GPU will configure for the AI and ML.

Also can you give an idea for the GPU bridging which may used to improve GPU to GPU communication speed to avoid PCI speed bottleneck?

Small correction we don’t currently support multi-GPU A100 and VMWARE with vCS. Support should be added in the next release of vGPU. We do support multi-GPU A100 with RHEL in the current release.

You can only have a VM on single host. So if your host has 2 GPUs then the max number of GPUs in the VM is 2.

You can use NVLINK between the two GPUs in the VM.

1 Like

How to buy nVlink for this card and where to buy. I am from Bangladesh. I was knocking to DELL but they are not offering vCS and nVlink Bridging. Can you help me on this please?

Hi Anis, can you please write to me at (sasi at nvidia.com). I will connect you with a reseller.