Using multiple graphics cards was a means to get more from your gaming PC by linking the two components together and taking advantage of the increased bandwidth. Unfortunately, seeing more than one massive GPU inside a desktop system is largely a thing of the past, and you'll likely only witness this scenario with specific niche applications, be it video editing, cryptocurrency mining, or enthusiast use. But I found a way to use my old GPUs by installing them inside my Proxmox server and deploying Frigate and OpenWeb UI.

Since it's difficult to use a multi-GPU setup for gaming with the lack of SLI and CrossFire support, I needed to look elsewhere to take advantage of some of the existing hardware I had from older systems. Thanks to AI upscaling and other technologies, GPUs are better-equipped to handle video output and rendering of even 4K content without much of an issue for frame rates, depending on the graphics card, of course. I had an Nvidia GeForce RTX 3060 Ti and GeForce RTX 4060 Ti lying around collecting dust, so I decided to put them to work.

While multi-GPU configurations may seem a little too niche in 2025, there's actually a good use case for it with Proxmox and a home lab. Should you have a service or two that can leverage the processing power and features of a dedicated GPU, this hypervisor platform can pass through multiple cards to various virtual machines and containers. This allowed me to create a multi-GPU setup for AI inference, video processing, and virtualized GPU passthrough, without even considering gaming.

Why multiple GPUs make sense

Create the ultimate consolidated home lab

I used to run a few servers inside a cabinet, each hosting a few services each to create a self-hosted empire for myself, my immediate family, and anyone else who managed to connect to our guest Wi-Fi. This was fine until I decided that forking out nearly 60 per month for our home lab was simply unacceptable. That's when I looked to consolidate everything into a single system. A powerful consumer-grade processor, such as an Intel Core Ultra 9 or AMD Ryzen 9, would do, but I decided to go all-out and use an AMD Ryzen Threadripper 9980X with a whopping 64 cores.

That's where the multiple full-size PCI slots come into play. You can do this on many consumer boards, too, but for Threadripper, I could throw in both the 3060 Ti and 4060 Ti and use them at full speeds, thanks to the sheer number of PCIe lanes offered by the CPU. Why would you wish to do this, however? It all comes down to what you plan to use Proxmox to run. I've got the hypervisor running Jellyfin, Frigate, and OpenWeb UI, and all three can utilize a GPU (or two!) for improved performance.

Running large language models (LLMs) locally, such as Llama or Mistral, requires significant VRAM and compute power. The RTX 4060 Ti, with its Tensor Cores and 16 GB of memory, is well-equipped to handle optimized models with great results. I can fire up qwen3:14b-q4 with just shy of 30 tokens per second, which isn't too bad and is not far off what I could expect through paid services. And if I weren't running Frigate on the same system, I could deploy both GPUs for dedicated model inference.

But there are other valid uses for multiple GPUs. Tools such as DaVinci Resolve or HandBrake can leverage multiple dedicated cards. This can significantly help lower the time needed to render real-time effects, color grading, and batch rendering. I've sat through long wait times with even an AMD Ryzen 9950X struggling to push through rendering 4K video, so having a GPU or two available can really speed up this resource-intensive process. But if you don't work with video or have any creative use for a GPU, the home lab could be the perfect solution for getting more from older parts.

I don't have the necessary capital to spare for a $2,000 GPU, which would be able to make quick work of running LLMs and handling multiple feeds with detection through Frigate, which is where multiple cheaper cards can pick up most of the slack. Depending on which cards you opt to use, you could even save a little on running costs, especially when undervolting them to lower power draw. My RTX 4060 Ti and 3060 Ti don't draw too much power, especially since the former is for sporadically running LLMs and the 3060 Ti can be undervolted to lower its draw considerably for detection work.

And should one GPU fail, I have a second ready to go immediately, which is handy for running a self-hosted and custom network video recorder (NVR) solution. I wouldn't want to be the bearer of bad news to my accountant that a new $2,000 GPU replacement is sorely needed.

Multiple GPUs + Proxmox = ❤️

It's a match made in heaven

Getting GPU passthrough to work on Proxmox is surprisingly straightforward, no matter if you're attempting to get a card through to a hosted Linux container (LXC) or virtual machine (VM). Both types of isolated instances can be configured to use hardware assigned by the underlying hypervisor. Where things can get a little tricky is when working with failover and resource management, sharing a pool of GPUs with multiple instances. I decided to keep it simple. The RTX 4060 Ti is solely responsible for LLM inference, and the RTX 3060 Ti is dedicated to Frigate.

So long as you have a motherboard (and CPU) with enough slots and lanes to power two (or more) GPUs, you're already pretty much ready to go. All that Proxmox needs is for IOMMU to be enabled in BIOS/UEFI and the official prep package (pve-nvidia-vgpu-helper) to be installed to get the OS ready for installing the necessary Nvidia drivers. Though we rely on Nvidia's own proprietary code and installers for this task, the entire process is straightforward and takes a couple of minutes to set up. There's fantastic official documentation on getting Nvidia GPUs set up with Proxmox.

One of Proxmox's strengths is isolation, and having both GPUs assigned (and effectively reserved) for two instances ensures there are no conflicts or risks of anything migrating between containers. More GPUs can be added later, or if I feel like using the same GPUs for other workloads, I can manage the deployment accordingly. It's incredibly versatile, allowing me to effectively run everything on one system without needing much overhead. Everything simply works too. Frigate can leverage the RTX 3060 Ti, and Ollama has no trouble filling the entire RTX 4060 Ti memory.

Make the most of old hardware

If you've built a PC or two, you may have a driver, motherboard, CPU, and GPU lying around collecting dust. Using Proxmox and some imagination, you can create quite the self-hosting platform with plenty of power to run your own AI chatbots, create a capable NVR security system for your home, and even automate everything on the IoT network. GPUs can really make a difference for the more demanding tasks, and so long as your PSU allows it, you can use more than one with Proxmox. I've found it incredibly useful.