Skip to content

GPU - H200🔗

Here the process of using H200 under Proxmox 9.1 is described. The used server is a GigaByte G494-ZB1-AAP2 that supports up to eight H200.

The GPUs should be passed through to different VMs. Within a VM it is planed to use MIG to run several containers in parallel to utilize parts of the GPU.

Prepare for Proxmox and PCI passthrough🔗

The following settings in BIOS are recommended:

Multi-Instance GPU (MIG)🔗

Enable MIG:

nvidia-smi -i 0 -mig 1

Disable MIG:

nvidia-smi -i 0 -mig 0

List possible MIG configurations:

nvidia-smi mig -i 0 -lgip

Create 4 instances of profile 15 (1g.35gb):

nvidia-smi mig -i 0 -cgi 15,15,15,15 -C

List existing MIG devices:

nvidia-smi -L

Run a CUDA-based version on specific MIG instances by using the ids starting with MIG- for CUDA_VISIBLE_DEVICES.

Delete the MIG devices shown in nvidia-smi:

nvidia-smi mig -i 0 -dci -ci <cid> -gi <gig>

Delete the MIG instances shown with nvidia-smi mig -i 0 -lgi:

nvidia-smi mig -i 0 -dgi