Reddit用户计划用24GB RTX 3090搭建本地AI服务器
一名Reddit用户称,其以1,100美元购入了一款配备24GB显存的EVGA RTX 3090 FTW3,以扩展容量超过仅有12GB显存的RTX 4070所能承载的本地AI工作负载。他计划将显卡安装在无界面的Ubuntu Server 26.04 LTS系统中,通过局域网为Ling-3.0-tiny、Omnivoice、Parakeet、代码推理及其他家庭实验室服务提供后端。帖子征求有关长上下文代码模型、按需切换模型,以及在没有EVGA Precision X1时管理FTW3在Linux系统下功耗与散热的建议。卖家关于显卡使用时间短、未挖矿且未用于AI计算的说法未获独立验证。
这一计划展示了一个具体案例:将本地AI推理任务从配备12GB显存的游戏显卡迁移至拥有24GB显存、具备24GB显存的独立服务器,同时探索模型服务与硬件管理方案。
Hey all,
I just picked up an EVGA RTX 3090 FTW3 (24GB) for $1,100 from a friend who is switching to an RTX 5070 Ti for pure gaming. It has low hours on it: no mining, no AI compute, just casual desktop use. He bought it for the prestige and never really pushed it.
Right now, everything local runs directly on my daily gaming rig with a single RTX 4070 (12GB):
- LLM / assistant: Ling-3.0-tiny
- Voice (STT/TTS): Omnivoice (Design + Cloning) plus Parakeet
12GB does not stretch far across all of that once I want a usable context window. On top of that, killing background inference pipelines every time I launch a game got old fast.
So the 3090 is going into my server box instead:
#EVGA#NVIDIA#RTX 3090 FTW3#RTX 5070 Ti#RTX 4070#Ling-3.0-tiny#Omnivoice#Parakeet#Ubuntu Server 26.04