The Minisforum MS-A2 has been my go to platform for running an optimized and fully functional VMware Cloud Foundation (VCF) 9.0 lab environment. From vSAN ESA to NSX VPC to vSphere Kubernetes Service to VCF Automation to Data Services Manager, it runs it all!
Over the past couple of weeks, I had also been experimenting with running VMware Private AI Services (PAIS) using the MS-A2. For those with an eye for detail, may have noticed that the NVIDIA GPU requirement was actually being satisfied by an ASUS NUC 14 Performance, which includes an NVIDIA GeForce RTX 4070 (8GB VRAM). The ASUS NUC was purely for validation and prototyping purposes to ensure that I could actually run PAIS before investing in a more capable and certainly more expensive NVIDIA GPU 🤑
One of the lesser known capabilities of the MS-A2 is the ability to add a half-height PCIe device and from the very beginning when I had first heard about the MS-A2, my plan and hope was to be able to add an NVIDIA GPU to the system!
With the compact form factor of the MS-A2, you are definitely limited by the types of PCIe devices that can be used and thats when I came across custom GPU coolers from n3rdware. Both the Minisforum MS-01 (Intel variant) and the MS-A2 can support NVIDIA RTX A2000, RTX 2000 Ada and RTX 4000 Ada.
I opted for the most capable GPU for running AI workloads and went with an RTX 4000 SFF Ada and with the largest amount of VRAM which was 20GB, definitely not a cheap purchase! 😅
Build-of-Materials (BOM):
- RTX 4000 SFF Ada with 20GB VRAM (VCNRTX4000ADALP-PB)
- RTX 4000 SFF Ada custom cooler
- Arctic Silver 5 Thermal Paste
Note: If you decide to buy a different GPU, just make sure you purchase a compatible cooler and that the GPU is matching what is supported since some of the online listings, its not always clear if it is single-slot or full desktop card. This is also why I listed the specific SKU for the RTX 4000 SFF Ada w/20GB VRAM. If you are interested in the other GPU cooler options, I have links to respective options for RTX A2000 & RTX 2000
The installation was super simple, you can either view the instructions from n3rdware or if you prefer watching a video, which I found quite informative in properly taking apart the GPU.
Here is a picture comparing the stock GPU shroud and the n3rdware custom cooler
Here is a picture of the raw GPU card, it is pretty insane how much of the GPU mass is actually made up of the heatsink, fan and shroud
Here is a picture of the final product, super slick looking and fits perfectly in the MS-A2!
😅 The moment of truth was to power on the MS-A2 and ensuring nothing fried and that the GPU fan was running and ESXi was able to properly detect the NVIDIA GPU! 🙌
how are use this with vgpu on esxi ?
I’m not, it’s used in passthru
Hey William,
I’m attempting something very similar on a Minisforum MS-02 (Ultra) running ESXi 8.0 U3 (build ESXi-8.0U3i-25205845) with an RTX 4000 SFF Ada (AD104GL).
Host detects the GPU fine and passthrough is enabled (both 02:00.0 + 02:00.1). Fresh VM boots perfectly without the GPU, but as soon as I attach it, power-on fails around ~88%.
From vmkernel.log:
“Dev 0000:02:00.0 is unresponsive after reset”
“Reset for device failed with Failure”
VM log ends with:
“Failed to find a suitable device for pciPassthru0”
I’ve tried:
* Fresh ESXi install + new VM
* 64-bit MMIO enabled (64GB)
* Full memory reservation
* pciPassthru.disableFLR = "TRUE"
* Different reset methods (bus/link)
* Passing both GPU + audio functions
Curious if you ran into any reset quirks on the MS-A2, or if your setup only worked reliably after a cold boot?
Appreciate any insight, feels like I’m very close but hitting a platform-level reset limitation.
Thanks!
Hi James,
I've not, as the blog post demonstrates this particular dGPU works perfectly fine in passthrough mode. The only difference between the two systems is that one is Intel-based and the other is AMD, so not sure if there's some quirk there ... If you can provide a vm-support bundle (provide a link to bundle), I can see if Engr sees anything based on your setup.
Hi William,
Really appreciate you taking the time to respond here 🙏
Quick update on my side, I was actually able to get this working reliably on ESXi 8 on the MS-02.
It turned out not to be a reset limitation at the hardware level, but an **MSI interrupt translation issue** on ESXi with this GPU.
The key fix was disabling MSI on the passthrough devices in the VMX:
```
pciPassthru0.msiEnabled = "FALSE"
pciPassthru1.msiEnabled = "FALSE"
```
After applying that (along with cleaning up the default NVIDIA wildcard in `passthru.map` and disabling PCIe hotplug), the behavior changed completely:
* VM powers on cleanly
* No more “unresponsive after reset”
* Reset operations now complete successfully
* VM restarts work
* Host reboot + VM start works consistently
vmkernel now shows clean:
“Reset device … done” and “Attached to device …” instead of the previous failure loop.
So it looks like the card itself behaves fine, but the MSI translation path was causing the failure on this Intel-based platform.
Your comment about Intel vs AMD might still be relevant here, as it seems like the interrupt handling path differs slightly.
Thanks again for the offer on the vm-support bundle, happy to provide one if it’s still useful for Engr, but at least there’s now a reproducible working configuration on this hardware.
Appreciate all the work you’ve put into documenting these setups 👍