Being able to reuse old components instead of pawning them for a few measly bucks has been a highlight of my home experience, especially with the RAM apocalypse causing PC hardware prices to spiral out of control. You see, contrary to what many newbies imagine, outdated hardware and dinosaur machines are enough to handle a few virtual guests.
This also applies to individual components, as you can salvage anything from old processors to used hard drives and put them to good use in a temporary server node. But if I had to pick one outdated peripheral that I absolutely refuse to part with, it would be my 10-year-old GTX 1080. Although Nvidia will end driver support for the Pascal series in December 2025, I plan to use this GPU in my home lab until it’s literally dead. And not because I paid a pretty penny for it as a broke teenager, but this beast of a GPU is still good enough for my LLM hosting and ML-backed tasks.
I primarily use a GTX 1080 to run native LLMs
With the MoE unloaded, it can run 26B models at decent token production rates
Before I went down the big language model rabbit hole, I was convinced that you needed an ungodly amount of VRAM to handle models that didn’t hallucinate like drunken lunatics when you asked them something remotely complex. However, the Mixture-of-Experts LLMs made me realize that my elderly gaming companion can still run high-spec models with “just” 8GB of VRAM and zero Tensor cores.
Thanks to a feature called MoE offloading, I can safely push expert weights to my computer’s CPU and RAM, leaving the heavy focus blocks on my Pascal cycle card. And because TN models enable multiple experts per token, I don’t have to worry about performance penalties when offloading the expert relay network to the slower components of my server.
For reference, I hooked a GTX 1080 into a Proxmox LXC armed with the llama.cpp packages, and as long as I allocated 24GB+ RAM to the container, my dinosaur-looking GPU could fit. Something big like Gemma-4-26B-A4B. Of course, it’s not as fast as the new RTX 50 series cards, but running the Gemma-4 at 14 tokens/second makes it a great addition to my home lab. So far I have combined it with Paperless-GPT, Paperless AI, Open the notebookHome Assistant (with the right plugins, of course), WinkKarakeep and a handful of other apps can take advantage of Gemma 4’s superior reasoning skills, and as long as I stay within the 150,000-token limit for the context window, I have no trouble responding to output requests from the arsenal I have on my GTX 1080.
But it can handle GPU-heavy workloads in certain FOSS applications
Video encoding and ML-accelerated tasks are still fair game for this Pascal card
Although I mainly use the GTX 1080 with llama-server LXC, it is also useful for other services that can directly use its processing capabilities. I have an Immich LXC responsible for handling all my images and screenshots, and it uses my decade-old GPU for hardware-accelerated facial recognition and intelligent search tasks.
Before moving my Frigate stack to a Raspberry Pi 5 and its AI suite, I configured the Pascal GPU to power object detection, pattern extraction, and face recognition tasks for all my security cameras. Likewise, I can now use Intel Quick Sync on my N100 system on Jellyfin (just so my GPU can focus on LLM-powered tasks), but I’ve long used the GTX 1080 to handle video encoding tasks on my media server.
Heck, I occasionally use it to power a remote gaming VM
It can run the vast majority of my Steam library at respectable frame rates
So far, I’ve only covered use cases where the GPU is integrated with my containerized applications. However, I sometimes temporarily disable GPU passthrough for my LXCs and instead pass it to a remote gaming virtual machine that streams Steam games to my smartphone using Parsec. Although it may seem a bit complicated, giving virtual machines access to the graphics card is fairly straightforward, and I keep most of the configuration files ready on my remote machine just in case I want to switch back to an LXC switch setup.
As strange as it sounds, this 10-year-old graphics card can still hold its own against modern titles. At 1080p, it can handle anything Hell 2, Sunless Skiesand other indie titles Cyberpunk 2077 (with medium-high graphics settings, of course), The Witcher 3, Dead Space (Remake), Man has no Heavenand other graphically intensive games. The only titles it encountered some drawbacks were multiplayer games with core-level anti-cheats (although Helldivers 2 is an exception) that refuses to load, as well as Chiaroscuro: Expedition 33here I had to lower the resolution a bit to get closer to 60 FPS. Otherwise, I have no problem streaming most of my Steam library to my smartphone at 1080p.
I plan to use the GPU until it finally breaks
As if my current home server experiences weren’t enough, I’ve subjected the GTX 1080 to even wilder projects in the past. Before upgrading to an RTX 3080 Ti, I armed it with a NZXT Kraken G12 mounting bracket + X52 liquid cooler combo to keep it from melting under my overclocking runs. For a 10-year-old graphics card (and considering all the pointless experiments I’ve done on it), it’s holding up pretty well, and given how useful it still is for my LLM hosting tasks, I don’t plan on retiring it any time soon.
