GeForce RTX 3090 (used)
The budget ticket to 24GB of CUDA. The local AI community's favorite used card for a reason.
Launch MSRP $1,499 (2020). Street prices move daily; the links open live listings. Amazon links may earn us a commission.
Will it run your models?
VRAM meterLocal AI fit: 32B-class models. Rule of thumb for 4-bit quantized models; context length and other apps eat into the same memory.
Why you buy it
- Cheapest 24GB CUDA card on the market
- NVLink support: two cards pool to 48GB
- Deep community knowledge for AI setups
Why you hesitate
- 2020-era gaming performance
- Many units lived hard lives in mining or render farms
- 350W of five-year-old silicon
The editor's take
Two generations old and still the cheapest reliable path to 24GB with CUDA, which is why AI hobbyists keep buying every one that hits the market. Gaming performance is around upper-midrange today, but for model inference the memory is what matters and the memory is why you are here. NVLink pairs two of them for 48GB, a trick nothing newer at this price can do.