Nvidia GeForce RTX 3090
Most local AI builds use one, for the 24 GB, 936 GB/s and CUDA. It draws 350 W under load, most are three slots thick, and they take two or three 8-pin leads. Used ones now sell above the $1,499 it launched at.
$1,500–1,700guide price, used
24GBmemory
936GB/smemory bandwidth
$91power a year at $0.18/kWh
What it runs
| 8B model, 4-bit | Fits on one card | |
|---|---|---|
| 14B model, 4-bit | Fits on one card | |
| 32B model, 4-bit | Fits on one card | |
| 70B model, 4-bit | Needs two of them |
Speed for local models follows memory bandwidth more than anything else, as long as the model fits in memory.
Specs
| Memory | 24 GB, 936 GB/s |
|---|---|
| Power | 25 W idle, 350 W under load |
| Power leads | 2 × 8-pin PCIe |
| Size | 3 slots, full height |
| Cooling | Own fans |
| Video encoder | Nvidia NVENC (Turing-class) and AV1 decode |
Listings on eBay
Live eBay listings for this model will show here.
Guide prices are used asking prices on eBay US checked 15 November 2026. Prices for cards with lots of memory have risen quickly this year.