I am looking for a way to share the same Nvidia GPU to multiple pods in Kubernetes.
As far as I know, it is not possible to do it directly, so I am looking for alternatives to do that.
With Nvidia vGPU technology, we could split a GPU into different vGPUs so that many services could use the same physical card. As Kubernetes is not a hypervisor, and the driver could no be used in this scenario, our idea is to put a hypervisor (KVM) between the bare-metal and the Kubernetes cluster, so K8s will be deployed in VMs. We are wondering if we can assign multiple vGPU created by the Nvidia driver to each VM so that we can map every vGPU to a specific pod of the Kubernetes cluster.
Is that possible? Is there any other way to do that?
Thanks in advance!