Nutanix Enterprise AI supports the NVIDIA GPU Operator when it is configured to use which GPU- virtualization mode?

Nutanix Enterprise AI solutions, which utilize the Nutanix Kubernetes Engine (NKE), require the NVIDIA GPU Operator to be deployed on Kubernetes worker nodes. For this configuration, Nutanix mandates the use of GPU passthrough. This method dedicates an entire physical GPU directly to a single worker node virtual machine. This direct assignment ensures maximum performance and compatibility for the NVIDIA GPU Operator, which then manages the GPU resource for containerized AI/ML workloads within the Kubernetes cluster.

A. GPU paravirtualization: This is not the virtualization technology specified or used by Nutanix AHV for high-performance NVIDIA GPU access in AI workloads.
B. SR-IOV: While AHV supports SR-IOV for networking, the prescribed and supported method for dedicating GPUs to NKE worker nodes for Nutanix AI is passthrough.
C. vGPU MIG: While Nutanix AHV supports NVIDIA vGPU and Multi-Instance GPU (MIG), the standard architecture for Nutanix Enterprise AI requires full GPU passthrough to the worker VM.

1. Nutanix Validated Design, AI-Ready Infrastructure with Nutanix: In the "Solution Design" chapter, section "Nutanix GPT-in-a-Box," it states, "The solution uses GPU passthrough to dedicate physical GPUs to the Kubernetes worker node VMs." (Reference: NVD-2463, Page 11).

2. Nutanix Kubernetes Engine Guide v2.9: In the "Enabling GPUs for NKE Clusters" section, under "Prerequisites," the documentation notes that for using the NVIDIA GPU Operator, you must first configure the GPU for passthrough on the worker node VMs. It states, "Add a GPU to each of the VMs that you want to use as workers in the GPU-enabled node pool. For instructions, see Enabling GPU Passthrough in a VM in the AHV Administration Guide." (Reference: Nutanix Kubernetes Engine Guide, Chapter: Enabling GPUs for NKE Clusters).

3. Nutanix University, Nutanix AI Configuration: Course materials for Nutanix AI deployment explicitly detail the process of configuring AHV nodes and VMs. The labs and lectures specify that for the Kubernetes worker nodes that will run AI workloads, the GPU must be attached using the passthrough (direct I/O) method before deploying the NVIDIA GPU Operator.