While building a GPU-enabled RDSH farm on Omnissa Horizon, provisioning failed on every single machine. vCenter looked healthy, the GPUs were empty and every host advertised the vGPU profile – yet Horizon insisted there was nowhere to put the VMs. This post walks through the troubleshooting, a couple of red herrings worth knowing about, and the actual fix.
The symptom
The farm went straight to an error state, and provisioning was disabled (because Stop Provisioning on Error is enabled):
Error during Provisioning Cloning of VM rdsh-16a-2 has failed: Fault type is UNKNOWN_FAULT_FATAL -No GPU capable host is available for provisioning rdsh-16a-2 with profile nvidia_l40s-16a.Please refer to this KB - https://kb.omnissa.com/s/article/59271 for more details.
The Events view showed Provisioning error occurred for Machine …: Resync operation failed for all machines in the farm.
The environment
- 3 × HPE DL380 Gen11, ESXi 8.0 U3, one NVIDIA L40S per host
- Host graphics: Shared Direct, vGPU mode Same Size, policy Spread VMs across GPUs
- Horizon instant clone RDSH farm, golden image with an nvidia_l40s-16a vGPU
- Two Connection Servers
Red herring #1 – the profile is there
The first check was obvious: does vSphere actually offer the profile? It does. Get-VgpuProfiles.ps1 (PowerCLI, on GitHub) lists the vGPU profiles each host offers together with the physical GPUs and their graphics mode, and every host listed the full L40S set (A, B and Q series). The Connection Server debug log even confirms it:
[InventoryValidationHelper] Successfully validated vGPU type nvidia_l40s-8q on HostOrCluster cluster01
So the profile is valid in vCenter. That turns out to be only half of what Horizon checks.
Red herring #2 – the vGPU profile lock
The second theory was GPU placement. It did not apply here, but it is a very common cause of the exact same error, so it is worth understanding:
- In Same Size mode, a time-sliced GPU is locked to the profile of the first vGPU VM powered on. Until all vGPU VMs on that GPU are powered off, it only accepts that exact profile.
- Mixed Size mode (vSphere 8.0 U3 and later, Ampere/Ada GPUs) allows different sizes on the same GPU – but only within the same series. An -8a (vApps) and an -8q (vWS) VM can never share a physical GPU.
- With one GPU per host, a running -8a farm spread across all hosts leaves zero hosts for a -8q farm – and Horizon reports “No GPU capable host”.
The quickest way to see what is really running on a GPU is nvidia-smi vgpu over SSH on the host. In this case the GPUs were empty, and the error persisted – so the problem had to be on the Horizon side.
Finding the cause in the Connection Server debug log
The lines that appear around the error in the admin console are just Horizon writing the error to the farm object in LDAP. The interesting part is what happens in the provisioning thread right before the error. To dig that out I wrote Get-HorizonGpuDebug.ps1, available on GitHub. It has two modes: a pattern search across the debug logs on one or more Connection Servers within a UTC time window (with context lines, exported to CSV), and -TraceError, which finds the first “No GPU capable host” error and prints the preceding lines from the same thread – usually where the placement decision is logged.
# Trace back from the first error (run on the CS that logged it).\Get-HorizonGpuDebug.ps1 -TraceError# Same, remotely against all Connection Servers (WinRM).\Get-HorizonGpuDebug.ps1 -TraceError -ConnectionServers cs01.domain.local, cs02.domain.local# Pattern search in a UTC time window, results to horizon-gpu-debug.csv.\Get-HorizonGpuDebug.ps1 -From '2026-11-19T11:15' -To '2026-11-19T11:25'
Start with -TraceError on the Connection Server that logged the provisioning error (check the Source column in Events). The output pointed straight at the problem:
[TaskStateDo] {HIC} Completed 1 shares for task ... CreateSingleCloneWrapperTask ...[CPHost] {HIC} VGPU_PROFILE is null or unknown.[CPHost] {HIC} VGPU_PROFILE is null or unknown.[CPHost] {HIC} VGPU_PROFILE is null or unknown.[CPHost] {HIC} VGPU_PROFILE is null or unknown.[CPHost] {HIC} VGPU_PROFILE is null or unknown.[Task] {HIC} Unexpected error executing ExecuteMultipleTasksAction ... No GPU capable host is available ...
One VGPU_PROFILE is null or unknown per host evaluated. Horizon never even asked vCenter to clone – it rejected every host internally.
Root cause – graphic-profiles.properties
The Connection Server keeps its own lookup table of known vGPU profiles and how many instances fit on a physical GPU: graphic-profiles.properties. If a profile is not listed there, Horizon cannot calculate host capacity and treats every host as not GPU capable – regardless of what vCenter reports. This is what KB 59271 refers to. In this installation the file had no nvidia_l40s entries at all.
The fix
Add the L40S profiles on every Connection Server, one server at a time. The easiest way is Add-HorizonVgpuProfiles.ps1, available on GitHub. Run it as Administrator on the Connection Server: it finds the file, backs it up, appends only the profiles that are missing (so it is safe to run again), and restarts the Connection Server service after asking you to confirm:
# Preview what would be added, change nothing.\Add-HorizonVgpuProfiles.ps1 -WhatIf# Add the missing L40S profiles, then restart wsbroker (asks first - drain the CS from the load balancer).\Add-HorizonVgpuProfiles.ps1# Update the file only, restart later.\Add-HorizonVgpuProfiles.ps1 -NoRestart
If you would rather do it by hand, this is essentially what the script does (note that running this manual version twice adds the block twice):
# Locate the file (path depends on Horizon version / branding)$conf = 'C:\Program Files\Omnissa\Horizon\Server\broker\conf\graphic-profiles.properties', 'C:\Program Files\VMware\VMware View\Server\broker\conf\graphic-profiles.properties' | Where-Object { Test-Path $_ } | Select-Object -First 1# Already there?Select-String -Path $conf -Pattern 'l40'# Back up and appendCopy-Item $conf "$conf.bak-$(Get-Date -f yyyyMMdd-HHmm)"@'# NVIDIA L40Snvidia_l40s-48q=1nvidia_l40s-24q=2nvidia_l40s-16q=3nvidia_l40s-12q=4nvidia_l40s-8q=6nvidia_l40s-6q=8nvidia_l40s-4q=12nvidia_l40s-3q=16nvidia_l40s-2q=24nvidia_l40s-1q=32nvidia_l40s-48a=1nvidia_l40s-24a=2nvidia_l40s-16a=3nvidia_l40s-12a=4nvidia_l40s-8a=6nvidia_l40s-6a=8nvidia_l40s-4a=12nvidia_l40s-3a=16nvidia_l40s-2a=24nvidia_l40s-1a=32nvidia_l40s-3b=16nvidia_l40s-2b=24nvidia_l40s-1b=32'@ | Add-Content -Path $conf -Encoding ASCII# Restart the Connection Server service (drain it from the load balancer first)Restart-Service wsbroker -Force
The format is profile-name=max instances per physical GPU, and the profile name is case-sensitive – it must match what vCenter reports. The values above follow NVIDIA’s limits for the 48 GB L40S; verify them against the NVIDIA vGPU user guide for your driver release.
When all Connection Servers are done, go to the farm and choose Maintain → Enable Provisioning. The machines were provisioned immediately, and the CPHost errors disappeared from the log.
Watch out during upgrades
The file lives in the Connection Server installation directory. Check it after every Connection Server upgrade – if the upgrade replaces it and the new build still does not include the L40S, the farm will break again at the next provisioning or push-image. Simply run Add-HorizonVgpuProfiles.ps1 again after each upgrade: it only adds what is missing, and changes nothing if the profiles are still there. It is also a good candidate for a post-upgrade step in your automation.
Bonus – carving up one L40S per host for RDSH
Since the troubleshooting forced a rethink of the layout anyway, here is what we ended up with for roughly 40 users (about 20 concurrent) on three hosts with one L40S each:
- One series and one size everywhere: nvidia_l40s-16a, Same Size mode. No profile lock surprises.
- 4 RDSH servers, 8 vCPU / 32 GB (fully reserved – required for vGPU), max 10 sessions each.
- 6 slots, 4 used: two vGPU slots per GPU at 16 GB. The two free slots let Horizon rebuild the farm after a host failure and give push-image room to work.
- Minimum ready machines during maintenance = 2, so a push-image never takes the whole farm down.
Takeaways
- “No GPU capable host is available” has (at least) two very different causes: a real placement problem in vSphere (profile lock, series mismatch, no free slots), or a Connection Server that does not know the profile.
- A successful validated vGPU type line does not mean Horizon can place the VM.
- Search the CS debug log for VGPU_PROFILE is null or unknown. If it is there, the fix is graphic-profiles.properties.
- If vCenter never shows a clone task, the rejection happened inside Horizon.