Hello,
We are currently developing a thor based system for a payload data processing computer. Our goal is to deploy multiple workloads (small models for inference, compression algorithms, …) from different entities at the same time in a virtualized environment.
One of the major concerns we have is how we forward the MIG instances to our hypervisor given that we don’t have the classical PCIe forward mechanisme as the thor GPU is not an external peripheral.
I would like to gather your experience on using virtualization on the thor:
- Did somebody successfully deploy a hypervisor with multiple VMs on the thor?
- How can we forward the GPU to VMs given that we don’t use PCIe?
- NVIDA seems to offer the NVIDIA AI Enterprise software to configure vGPUs. Does this solve our problem here and is it possible to get this for the thor?
Thank you in advance for your inputs.