NOTE: This is not an elegant solution, as 1 GPU = 1 Container.
In my use case, I have an Ubuntu desktop for my testbed where Proxmox lives inside of Ubuntu and beside also Windows. This is a bit of a pain to get going, but after it’s set up my testbed is one computer. This is particularly useful for me, as during the summer it gets very hot here and I don’t need the full cluster for testing various containers on-the-fly without the need for redundancy (yet).
It’s sort of the blight of many Proxmox users at this time, I think, to scrap everything and start over time to time. Sure there are better ways to do this but sometimes it’s just faster to scrap and reset literally everything (just being honest).
Now if you’re down to one computer, that’s when it’s really nice to have lots of threads. For me I finally got around to switching to my workhorse - a Threadripper 1950X which was last year the mothership of my stack and this year is my main squeeze. I finally treated myself to an RTX 2060 12GB (not 6, mind!), so now my new primary has team red (RX 480 8GB) and team green Nvidia for that CUDA as well.
Why RTX 2060? Well, it’s the last Nvidia card to (technically) support SLi and it came with 12GB which manufacturers are skimping you on for newer cards.
I will probably get around to using MSI afterburner to cool things off next summer and keep this hardware lasting for as long as possible. So far I’ve completely amortized the Radeon card ($150 in 2016) and here’s to hoping EVGA is still the quality my computer shop has lived by. These are really my only two consumer cards that aren’t ancient, though some might say even this computer tech’s setup is getting quite dusty. It just goes to show you, new hardware is overrated (mostly), though if I wanted to go thin, nimble, and local I would opt for the newer hardware in that case.
Speaking of case, I’m rocking a CoolerMaster LAN party. It has proved very future-proof and relatively discrete, maybe even a good conversation piece.
I’ve not discussed it in depth, but most of my work is seasonal as with the seasons. Things really “heat up” around here in winter, while in summer I’m on the thinnest client I can limp along with, which unfortunately for me is still largely on the cloud. Hopefully by next summer local AI is easier to work with and more powerful.
You can run your full stack year round, but then you encounter the power usage competing your air conditioner with your space heater. Much better for me to only use what I need, all else equal. If I had all the money in the world, I’d throw money at both year-round, youuuu betcha.
For our Windows virtualized inside of Ubuntu latest setup, I’ve found Qemu/KVM is the most practical. I don’t much care for its UI nor default parameters revolving around GPU passthrough, but the rest I chose simply because my preferred Oracle solution was conflicting with Qemu/KVM, so Qemu/KVM won the day.
The alternative is using WSL inside of Windows. Now this is getting simpler to use here in 2025, but still has some trouble with network or port conflicts and my Visual Studio co-pilot really likes to spend tokens forgetting whether it’s in powershell or WSL which are tiring to override.
Actually right now I’m dual-booting 1) Windows with WSL 2.0 (Ubuntu) inside of it on one M.2. and 2) My newer instance which is Ubuntu Latest with both Proxmox and Windows installed on another M.2. It’s a good idea to have a Raid1 (Mirror) backup array and an external cold storage USB 3.1 hard drive or even tape drive (have you heard of such a thing?!) if you’re down to only one client.
Yes, it is possible to window yourself where you forget what environment you’re actually in if you’re not paying attention to your workflow. Careful consideration of the parameters of your preferred setup before you roll up your sleeves is a must - measure twice, cut once. Or just dual-boot and virtualize in each. What’s your fancy?
I do see why Threadripper is overpriced. It took me some haggling with an old customer and with Ebay to frankenstein my 1950X 32-fast-thread with an ebay board that came with an extra 1900X that is just waiting to be re-sold. But man, when I finally got it going, I have ample wiggle room for lot of cards including also a PCI-e M.2 SSD and with eight memory slots I can expand to 128GB for about average competition with new gear, not bad for old hardware.
I do really want to get my stack online. Part of the problem is my chosen environment is a cold basement and I’m kind of a lightweight. Factor the need for a jacket and a good attitude in the cold if you’ve chosen the right spot, I’ve often said.
As with most things, the computer goes at exactly one speed but the human’s gotta feed the monkey. Factor that in too, if you’re being honest.
