Nvidia DGX-2 SYSTEM User Manual page 68

NVMe drives as pass-through devices

To preserve the existing RAID configuration on the DGX-2 System and simplify the
process of reusing this resource if the server were ever to be reverted from KVM,
NVMe drives cannot be used as pass-through devices.
The DGX-2 KVM host cannot be used to run deep learning containers with GPUs.

NVIDIA GPUDirect

11.1.2 About the Guest GPU VM (Features and
Limitations)
Guest GPU VMs are based on an installed KVM image.

Guest GPU VM size and resources are based on the number of GPUs assigned

Once a GPU VM is created and resources assigned, reconfiguring the VM (adding or

removing GPUs, modifying other resource allocations) is not supported.
Access to the hardware is restricted from within the guest GPU VM such that

GPUs cannot be reset
●
GPU VBIOS cannot be updated
●
System firmware (including Mellanox IB) upgrade is not supported
●
There is no guest UEFI BIOS support
●
11.1.3 About nvidia-vm
Guest GPU VMs can be managed using the virsh (see https://linux.die.net/man/1/virsh)
program or using libvirt-based XML templates. For the NVIDIA KVM, NVIDIA has
taken the most common virsh options and configuration steps and incorporated them
into the tool nvidia-vm, provided with the DGX KVM package. nvidia-vm simplifies
the process of creating guest GPU VMs and allocating resources. In addition, you can
use nvidia-vm to modify default options to suit your needs for the VM and manage
VM images installed on the system.
To view the top-level help, enter the following.
nvidia-vm --help
You can view the man pages by entering the following from the DGX-2 KVM host.
man nvidia-vm
Details of basic commands are provided in the following sections.
DGX-2 System User Guide
is not supported on multi-GPU guest VMs over InfiniBand.
TM
Using DGX-2 System in KVM Mode
68