[slurm-users] Nvidia virtual GPU (vGPU) and Slurm?

[Timony, Mick]

Hi,

I've been considering new purchasing NVidia RTX6000 or the  RTX8000 NVidia GPU's to add to our existing GPU's partitons on our  Slurm cluster.

The RTX6000 has 24GB of on-board memory and the RTX8000 has 48GB, both of these are single-precision cards. Besides the additional 24GB of memory the RTX8000 supports something that NVidia calls virtual GPU (vGPU) the RTX6000 does not support vGPU. This allows one to carve up the RTX8000 to appear as multiple GPU's.

Are there any Slurm installations that have tried using or are using vGPU technology with these or other NVidia GPU's? Can you share you experiences using this technology?

Was it easy to configure?
Any issues getting it to work with Slurm?
How stable has the technology been for you?
Has it caused any the GPU's to be unstable or to crash more or less often?
Is there any overhead of using vGPU?

I would appreciate to hear any feedback to these questions and your thoughts and concerns about using vGPU with Slurm.

Kind regards

--
Mick Timony
Senior DevOps Engineer
Harvard Medical School
--

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.schedmd.com/pipermail/slurm-users/attachments/20200701/abb59fb6/attachment.htm>