llama.cpp CPU performance in the big 26
A user inquired about running llama.cpp on an old decommissioned office desktop as a backup LLM server, specifically considering qwen 9b, due to the high cost of new hardware. They currently use a 3090 with qwen 3.8 27b on their main server. After reviewing community replies, the user concluded that setting up a powerful enough LLM on such hardware would be too problematic. Instead, they plan to use a backup machine for non-LLM services.
This post highlights the practical challenges and cost-benefit considerations of deploying LLMs like qwen 9b on older, less powerful hardware, unlike discussions focused solely on high-end systems.
Time & source
Times shown in UTC
Display time zone: UTC
Local time zone unavailable; showing UTC.
PublishedOffset at this time: UTC+0Sep 12, 2026, 19:17 UTC
IngestedOffset at this time: UTC+0Sep 13, 2026, 15:01 UTC
- Published
- Sep 12, 2026, 19:17
- Ingested
- Sep 13, 2026, 15:01
- Source type
- Dev community
- Tier
- Community
- Source status
- Healthy
Tier is a per-source editorial setting, not a per-item score.
my main server has a 3090 with qwen 3.8 27b (crazy who woulda guessed) but i want to a have a backup failover sever with a somewhat useable LLM offsite (my parent's home). because infinite money glitch ai bubble has made hardware much harder to afford for us proles, was thinking of getting an old decommisioned office desktop (i.e. optiplex, elitedesk, etc) and having it boot up llama.cpp with qwen 9b or something if my server ever shits the bed.
how good is cpu performance on older mid tier cpus with ddr4 ram? or cpus in general? is it useable? what can i expect, and how can i set it up? is it even a good idea in the first place?
edit: thanks all. reading through the replies, seems like it's more of a headache than its worth. i'm already looking for a backup machine since this does NAS and other stuff but i think i'll just import all other services that aren't LLM into it, and not be beholden to trying to find something powerful enough for a useable LLM