Hi all,

I’ve been running a Proxmox server for simulation workloads. The idea is simple: either the Windows or the Linux VM runs (never both at once, I use a hookscript to enforce that), and they get as much CPU and RAM as possible. A TrueNAS VM runs permanently to provide shared storage via NFS.

The problem is with the Windows VM. As soon as it starts a heavy simulation, at some point the entire server freezes — no SSH, no web UI, no ping. I’ve had to hard reset it multiple times.

System

• Proxmox VE 8.4.0 (6.8.12-9-pve)

• AMD Ryzen Threadripper 7980X (64c/128t)

• ASUS Pro WS WRX90E-SAGE SE

• 512 GB DDR5 ECC (8× Kingston 64GB 5600MHz)

• Samsung 990 PRO 1TB (ZFS boot + 500 GB NFS export)

• Crucial P3 Plus 4TB

• GIGABYTE RTX 4070 Ti SUPER (passed to Windows or LINUX)

• Thermaltake ToughPower PF3 1050W

• Case: be quiet! Silent Base 802

Proxmox is installed on a ZFS mirror (RAID1) using two Samsung 990 PRO SSDs. A 500 GB partition from this pool is shared via NFS directly from the Proxmox host. The TrueNAS VM runs separately and shares the larger 4TB SSD over the network.

VM setup

Windows VM

• 400 GB RAM (no ballooning)

• 56 cores (1 socket)

• CPU: host

• GPU passthrough enabled

• Disk: local-zfs

Linux VM

• Same concept, not running at the same time

TrueNAS VM

• 16 GB RAM

• Always running (serves NFS)

• Disk is on rpool (to avoid ZFS-on-ZFS)

What I’ve tried

• Reduced RAM to 200 GB, then 100 GB → still crashes

• Disabled ballooning

• Checked logs (dmesg, journalctl) → no OOM, no PCI/GPU errors

• Swap file (16 GB) added

• Host is thermally fine post-crash

• NUMA is enabled

• System is stable under bare-metal stress

What I’m wondering

Could GPU passthrough still cause issues even if it works at first? Are there known problems with high-core AMD setups in Proxmox 8.x? Would switching away from local-zfs help? Is 56 cores + 400 GB just too much for a single VM?

Appreciate any pointers — happy to post qm config or logs if useful.

Hi all,

I tring to build a proxmox homelab server to run some VMs. I need at least 32 cores as I have to test some SIEMs, create windows AD forest and get log correlations and test some advanced incident response tools.

I have already looked into AMD's Epyc processors, even if Max TDP and yearly running cost is lower on Epyc I think it's not the best buy for my usecase.

The system won't be running 24/7, I will only use it when I have do perform above tests and valuations.

I was thinking of the following setup:

AMD Threadripper 3970X 3.7 GHz 32-Core Processor  
Noctua NH-U9 TR4-SP3 Dual Tower Cooler  
MSI TRX40 PRO 10G ATX sTRX4 Motherboard  
2x Samsung 970 EVO Plus SSD 1 TB  
Corsair RM750x 750W  
8x Crucial Ballistix Sport 3200 MHz 16GB  
Quadro P400 2 GB Video Card  
Dark Base Pro 900 rev.2 Case  

Feedback and input are welcome, particularly from others who have built ThreadRipper for proxmox.

Did you encounter any issues for a similar built?How is the quadro p400 pass-through on proxmox or any other relevant information for the evaluation of the project?

Hello,

We are planning on building a shared rig for machine learning. Of course, we are cash-strapped, so we are trying to make the best of it. I have done my share of proxmoxing with 6.4 on a Dell Precision 7810 bi-Xeon a short while ago with Quadro boards and the fun of passing them to the host.

On that new machine, the load would be basically 2 VMs: one of a CAD Software (Solidworks) and another one for a linux for machine learning (Ubuntu). The linux box would have one GPU to start, and grow to two if/when we max it out.

So the motherboard would be running at least 2 or max 3 GPUs, which will be blacklisted in the host and assigned beforehand to each VM.

Xeons are too expensive, and I am looking at getting either a 59xx threadripper or a 7950x Ryzen (aka bleeding edge). Here are my questions:

1. I seem to recall that some logic boards / threadripper or ryzen combos where hit or misses with proxmox. Is this still the case? Is there an up to date list like 'don't touch that one with a 10ft pole'?

2. What is the support for NVidia RTX 30xx (Ti or plain)? I recall some of these would 'code 43' in Windows, so not so good?

Any hints welcome on the strategy etc...

Did Ryzen gave some of you some headaches (vs Intel)? Or is it working just fine?

Edit: (I'm thinking about the Ryzen 5 5600)

I recently purchased the components to build a new AI dedicated PC running on the WRX90e with a 5090 suprim liquid, I was curious if anyone has any suggestions for an AIO built for the sTR5 processor type. I've seen a couple out there but they were pretty bare, I wouldn't mind a nice fancy looking cooler. If I have to build one, I guess I could go that route, Any suggestions?

For anyone else looking, the two finalist I found were:
Gigabyte AORUS Waterforce X II 360
Corsair iCU H150i Elite LCD XT

I went with the AORUS.