GPU Hosts (NVIDIA vGPU and Passthrough) — Build Guide

This guide gets NVIDIA GPUs in ESX hosts ready for workloads on VCF 9.1: choosing how the GPU is shared, which licences that pulls in, checking the card against both Broadcom’s and NVIDIA’s support lists, getting the NVIDIA host driver into the cluster’s vSphere Lifecycle Manager image, and configuring the GPU devices on each host.

It stops where the platform is ready. What runs on top — deep learning VMs, VKS clusters with GPUs, Private AI Services, VCF Automation catalog items — is covered by Broadcom’s VMware Private AI Foundation with NVIDIA documentation.

Evidence level: sourced from Broadcom TechDocs (the VMware Private AI Foundation with NVIDIA 9.1 PDF, generated 2026-09-21, and the VCF 9.1 lifecycle pages), the Broadcom Compatibility Guide (queried 2026-10-16) and NVIDIA’s vGPU documentation (pages last updated 2026-09-29). Not lab- or field-verified in this repo yet. NVIDIA ships vGPU releases often and Broadcom updates the compatibility guide constantly, so re-check both lists (§3) on the day you order hardware.

The sequence, end to end

1. Pick the GPU mode: passthrough, vGPU time-slicing, or vGPU MIG
2. Line up the licences
3. Check the card on both lists
4. Pick the vGPU release and download it
5. Enable SR-IOV in the BIOS
6. Put the host driver into the cluster image — before the workload domain or cluster is created
7. Configure the GPU devices on each host
8. NVIDIA licensing for the guests (vGPU only)
9. Guest drivers
10. Validate

Steps 1–4 are planning: they belong with the hardware order, because the answers decide which card and which licences you buy.

1. Pick the GPU mode

Points that decide the choice:

• SR-IOV is required for both vGPU modes. PAIF 9.1: “SR-IOV is also required for both MIG and time-slicing modes of NVIDIA vGPU. For MIG, a vGPU is associated with a virtual function at boot time.”
• MIG and NVIDIA NIMs don’t mix. PAIF 9.1: “You cannot use the MIG sharing mode for NVIDIA GPUs with NVIDIA NIMs.” If the AI team plans to run NIMs, plan time-slicing or passthrough on those GPUs.
• vGPU VMs can vMotion once a vCenter advanced setting is on (§7.4).
• One mode per GPU device, set per host. Switching a host from MIG to passthrough means turning MIG off and removing the host driver (§7.3).

2. Line up the licences

Up to three licence layers apply. Name an owner for each before the hardware arrives.

• NVIDIA’s evaluation. NVIDIA offers a free 90-day vGPU evaluation. It needs a corporate email address, and the 90 days start when the activation key is redeemed in the NVIDIA Licensing Portal. NVIDIA’s pages disagree on the concurrent-user cap (128 vs 500), so check the evaluation page when you register.

GPUs for VMs, or Private AI Foundation?

Steps 3–10 of this guide are the same either way. A Private AI Foundation implementation adds, on top of plain GPUs for VMs (PAIF 9.1, Deploy a Workload Domain):

• the add-on licence, assigned to the GPU workload domain’s vCenter — which only works once that vCenter is in a vCenter Linking group with the management vCenter;
• the workload domain’s vCenter and NSX Manager connected to the VCF Identity Broker (fleet SSO);
• at least 3 GPU hosts in the initial cluster;
• its own setup workflow for what runs on top (deep learning VMs, VKS clusters with GPUs, Private AI Services, VCF Automation catalog items), which on a disconnected (dark) site means staging every artifact locally. That part is 33-private-ai-foundation.md.

Plain vGPU or passthrough for ordinary VMs needs none of that: VCF plus, for vGPU, the NVIDIA licence.

• Entitlement lag. A new entitlement “might not appear on the NVIDIA Licensing Portal dashboard page until 24 business hours” after you first set your password (NVIDIA quick start). Register early, not on build day.

3. Check the card on both lists

A GPU has to be supported by Broadcom (for this ESX release, in this mode) and by NVIDIA (in the vGPU release you will install). The two lists are maintained separately and don’t always agree (§3.4).

3.1 Get the card’s exact IDs

A model name isn’t enough: the compatibility guide lists each card by its PCI VID / DID / SVID / SSID. On a host that already runs ESX:

esxcli hardware pci list

Find the entries with Vendor ID 0x10de (NVIDIA) and note the Device ID, SubVendor ID and SubDevice ID. Before the hosts exist, ask the server vendor for the same four values for the exact GPU part in the quote.

3.2 Broadcom Compatibility Guide

Search the Broadcom Compatibility Guide for your ESX release:

• GPUs program — filter Releases = ESXi 9.1, Partner = NVIDIA, and Supported Passthrough Mode = NVIDIA vGPU (or DirectPath for passthrough). Match on the four IDs from §3.1. Example: NVIDIA L40S (10de:26b9, SSID 18cf), listed for vGPU and DirectPath on ESXi 9.1.
• DirectPath GPUs and Accelerators program, Compute = AI/ML — the list Private AI Foundation points to: “verify in the Broadcom Compatibility Guide for GPUs and Accelerators for AI/ML Compute that the GPUs on your ESX hosts are supported for AI and machine learning use cases by Broadcom” (PAIF 9.1).
• AI Ready Node (Server + GPU) program — certified server + GPU combinations (ESXi 9.1, 9.0 and 8.0 U3 at the time of writing). Worth checking first when the server isn’t bought yet.

3.3 NVIDIA vGPU support matrix (vGPU only)

NVIDIA’s product support matrix for VMware vSphere lists the supported GPUs per vGPU release. Since vGPU 19.x it lists VCF 9.1 and VCF 9.0 as their own platform rows, separate from vSphere 8.0:

Footnotes worth knowing (20.x matrix):

• RTX PRO 6000 Blackwell Server Edition (incl. liquid-cooled) and RTX PRO 4500 Blackwell Server Edition need VCF 9.0.1 or later on the 9.0 line.
• 20.x dropped the Tesla M10, several Tesla V100 variants, and the Quadro RTX 6000 / RTX 8000. On VCF 9.1 those cards stay on 19.x at best.
• vGPU needs ESX licensed as vSphere Foundation (or vSphere Enterprise Plus), which VCF covers.
• 20.x adds MIG-backed vGPUs for graphics on vSphere 9 Update 1, and (20.2) multiple MIG-backed vGPUs on the same physical GPU on VCF 9.1 or later (20.2 release notes).

3.4 The two lists disagree

Captured 2026-10-16:

A card on Broadcom’s list is not proof that NVIDIA’s current vGPU release supports it. For an older card, find the newest vGPU branch that supports it and has a VCF 9.1 package, and confirm that NVIDIA still supports that branch.

4. Pick the vGPU release and download it

4.1 Which release

• Use a release that ships a “NVIDIA vGPU for VCF 9.1” package: 19.4 or later on the 19.x branch, or 20.x. Don’t use the vSphere 8.0 package on VCF 9.1.
• Host and guest drivers come from the same release. The host driver must be “compatible with your VCF version” (PAIF 9.1), and the guest driver must match the host driver. Private AI Foundation’s deep learning VM pins this too: the DLVM for 9.1 supports “NVAIE 7.x vGPU host driver (580.x.x) and NVAIE 8.0 vGPU host driver (595.58.02)”, and the 9.1.1 DLVM “NVAIE 8.0 and later versions of vGPU host drivers (595.x.x)” (PAIF 9.1). If you’ll run DLVMs, pick a host driver inside that window.
• For VKS clusters with GPUs, the NVIDIA GPU Operator must be compatible with the cluster’s Kubernetes version and with the host driver (PAIF 9.1).

4.2 Where to download

Download the offline bundle (ZIP) for the host driver: that’s what vSphere Lifecycle Manager imports (§6.1).

5. Enable SR-IOV in the BIOS

For vGPU (time-slicing and MIG): enable SR-IOV in each host’s BIOS/UEFI, following the server vendor’s documentation, and on the graphics devices using the VMware Host Client (PAIF 9.1). Do it during host preparation, alongside the rest of the BIOS baseline, so it isn’t a maintenance-mode round per host later.

6. Put the host driver into the cluster image

VCF lifecycle-manages ESX through vSphere Lifecycle Manager images, so the NVIDIA host driver goes into the image as a component — not hand-installed on hosts and left there. Private AI Foundation lists “A vSphere Lifecycle Manager image with the VIB file of the vGPU host manager driver” as a requirement, and installs the driver by “Add the VIB file of the driver to the vSphere Lifecycle image for the cluster and remediate the hosts” (PAIF 9.1).

Do this before creating the workload domain or cluster, so you can select the GPU image there.

Source conflict, dated: the PAIF requirements still say the image must be “available in SDDC Manager” (PAIF 9.1, generated 2026-09-21). The VCF 9.1 lifecycle page (last updated 2026-10-15) says you “create vSphere Lifecycle Manager images in the vSphere Client and then import them into a VCF Instance using VCF Operations” (Managing vLCM images). The SDDC Manager UI is deprecated in VCF 9; follow the VCF Operations path below.

6.1 Build the image in vCenter

In the vSphere Client of a vCenter that runs the target ESX version:

1. Import the NVIDIA offline bundle (ZIP) into the vSphere Lifecycle Manager depot. KB 425993 gives the path as Lifecycle Manager → Image Depot → Import Updates, and prefers a ZIP offline bundle over an ISO.
2. Create the image in the image library (Working with Image Library): the ESX version from the VCF BOM (Import a vLCM image), your vendor add-on and firmware as usual, and add the NVIDIA host driver as a component.

Mixed clusters (some hosts with GPUs, some without, or different GPU generations) can use a composite image: one default image plus up to four additional image definitions with the same base ESX version but their own vendor add-on, firmware and components, assigned to hosts manually or by BIOS vendor / model / OEM string / family rules (Managing vLCM images). Clusters whose hosts all carry the same GPUs don’t need this.

6.2 Import the image into the VCF Instance

TechDocs:

1. Log in to VCF Operations as an Administrator.
2. Build → Lifecycle, expand VCF Instances, select the instance.
3. Image Management → Import Image.
4. Import from a vCenter (pick the vCenter and the image) or Import from a file (JSON + ZIP, optional ISO, exported from the vSphere Client first).

6.3 Use the image

• New workload domain or cluster: select the GPU image in the creation wizard (23-workload-domain-creation.md, 24-cluster-creation.md). When you use the Private AI Foundation guided workflow, “select hosts whose NVIDIA vGPU state is Ready” (PAIF 9.1).
• Existing cluster: add the component to the cluster’s image and remediate the hosts (PAIF 9.1).
• Private AI Foundation minimum: “At least 3 GPU-enabled ESX hosts to include in the initial cluster of a workload domain” (PAIF 9.1).

Upgrades: a VCF upgrade moves ESX to a new image, so the NVIDIA component has to move with it. Check the vGPU release still supports the target VCF version (§3.3) and include a matching host driver in the target image before you start the ESX upgrade.

7. Configure the GPU devices on each host

All in the vSphere Client of the workload domain’s vCenter, per host (PAIF 9.1, Configure NVIDIA vGPU or GPU Passthrough for AI Workloads on the ESX Hosts). Each change needs the host in maintenance mode and a restart, so batch them per host.

7.1 vGPU: Shared Direct

1. Select the host → Configure → Hardware → Graphics.
2. Under Host Graphics, select a GPU device and click Edit.
3. Set Device Type = Shared Direct.
4. vGPU Mode (dual-purpose GPUs only): TechDocs, verbatim: “vGPU profiles on a single physical GPU can have different types or sizes. For example, you can run vGPU VM profiles with 16-GB compute for compute-intensive work like machine learning (C-type vGPU profile) and 4-GB graphics (Q-type vGPU profile) simultaneously on a single device. Activating this feature reduces the maximum number of instances of some profiles.”
5. Repeat for every GPU device, then maintenance mode → restart.

7.2 Passthrough: Fixed or Dynamic DirectPath

1. Select the host → Configure → Hardware → DirectPath.
2. Select a GPU device → Configure DirectPath Mode → Fixed/Dynamic DirectPath → Configure.
3. Repeat for every GPU device, then maintenance mode → restart.

7.3 MIG

• Enabling: MIG is enabled per host, on each host in the cluster, following NVIDIA’s MIG User Guide (PAIF 9.1).

• Switching a host from MIG back to passthrough (PAIF 9.1): maintenance mode; on the host, get the GPU IDs with nvidia-smi; turn MIG off per GPU and verify; reboot; then configure DirectPath (§7.2).

  nvidia-smi                      # list GPU IDs
  nvidia-smi -i <gpu_id> -mig 0   # MIG off for that GPU
  nvidia-smi -i <gpu_id>          # verify
  reboot

  Turn MIG off before removing the host driver. If the driver is already gone, TechDocs’ workaround is to run the nvidia-smi -i <GPU_ID> -mig 0 command from inside a VM that has nvidia-smi.

7.4 Allow vMotion of vGPU VMs

On the workload domain’s vCenter (PAIF 9.1):

1. Hosts and Clusters → select the vCenter instance.
2. Configure → Advanced Settings → Edit Settings.
3. Set vgpu.hotmigrate.enabled to Enabled → Save.

8. NVIDIA licensing for the guests

vGPU guests check out a licence from a service instance of the NVIDIA License System. Set it up before the first vGPU workload, because an unlicensed vGPU slows down. NVIDIA, verbatim: “the performance of the virtual GPU or physical GPU is degraded if the VM fails to obtain a license within 20 minutes”, and it degrades further after 24 hours (NVIDIA licensing intro).

• Which service instance: a cloud-hosted License Service instance on the NVIDIA Licensing Portal when the guests can reach it, or the on-premises Delegated License Service (DLS) virtual appliance on a disconnected (dark) site — “you host the NVIDIA licenses locally on a Delegated License Service (DLS) appliance”, deployable “in the same workload domain as the AI workloads or in the management domain” (PAIF 9.1).
• The licence server software is downloaded from the same Licensing Portal (§4.2).
• Each guest gets a client configuration token in the default folder (NVIDIA client licensing):
  • Windows: %SystemDrive%\Program Files\NVIDIA Corporation\vGPU Licensing\ClientConfigToken
  • Linux: /etc/nvidia/ClientConfigToken
• For a vGPU you don’t set a licence type: the software “selects the license type based on the vGPU type” (NVIDIA client licensing).

Bake the token into your templates (or let the deep learning VM / GPU Operator handle it, per the Private AI Foundation docs), and give the licence server an owner: it is a new piece of infrastructure, with its own backups and renewals.

9. Guest drivers

10. Validate

• Every GPU host is on the GPU image and compliant in vSphere Lifecycle Manager (no drift from the cluster image).
• nvidia-smi on each host lists every GPU (vGPU hosts; it ships with the host driver).
• Configure → Hardware → Graphics shows Shared Direct on every vGPU device; DirectPath shows the passthrough devices as configured.
• MIG mode matches the plan per GPU (nvidia-smi -i <gpu_id>).
• vgpu.hotmigrate.enabled is set on the workload domain vCenter, and a test vGPU VM vMotions between two GPU hosts.
• A test VM with a vGPU profile powers on, runs nvidia-smi in the guest, and gets a licence from the NVIDIA License System.
• GPU metrics appear in vCenter and VCF Operations.

Not covered here

• Advanced NIC passthrough (Enhanced DirectPath I/O and Uniform Passthrough for NVIDIA ConnectX-6 DX / ConnectX-7 / BlueField-3, GPUDirect RDMA). Those clusters are added by hand to a dedicated vDS and transport node profile; with such hosts as the first cluster, the workload domain must be created without an initial cluster (PAIF 9.1).
• Device groups (NVLink / NVSwitch GPUs presented as one unit) and HGX platforms.
• Everything above the hosts: VM classes, deep learning VMs, VKS GPU clusters, Private AI Services, VCF Automation catalog items, and the dark-site staging — see 33-private-ai-foundation.md.

References

Broadcom: VMware Private AI Foundation with NVIDIA 9.1 (PDF) · Private AI Foundation 9.x docs · Requirements for Deploying Private AI Foundation · Configure NVIDIA vGPU or GPU Passthrough on the ESX Hosts · Managing vSphere Lifecycle Manager Images for VCF (9.1) · Import a vSphere Lifecycle Manager Image to a VCF Instance (9.1) · vSphere 9.1 — Working with Image Library · KB 425993 · KB 367541 — Installing and configuring the NVIDIA VIB on ESXi · Broadcom Compatibility Guide · VCF 9.1 is an NVIDIA-Certified Hypervisor (VCF blog, 2026-08-27)

NVIDIA: vGPU support matrix — VMware vSphere (latest) · 19.x matrix · 18.x release notes (archived) · vGPU updates by release · Getting NVIDIA vGPU Software · vGPU evaluation · NVIDIA License System · Client licensing — introduction · Client licensing — configuring a licensed client · NVAIE VMware deployment — software · NGC — Virtual GPU Manager · MIG User Guide · Public driver downloads