#$400-500 AI GPU; Nvidia; United States; Used is fine

1 messages · Page 1 of 1 (latest)

fading bolt
#

thank you!

drifting wadi
crude tartan
#

3060 12G is better tbh. 2GB more vram. 3090 comes next, and 5090 for ultimate performance

oblique vale
crude tartan
#

Depends on what you’re doing. 3080 is fine until you want to run a model that’s more than 10GB vram. Then it literally will not work

drifting wadi
#

I was trying to stay in budget. Depends on OP’s needs

crude tartan
#

I’m saying you want the most vram possible, so a 3060 has 12, and is pretty cheap. You can get a pair if you want

oblique vale
crude tartan
#

that would be way slower

oblique vale
crude tartan
#

how? if it has to wait anyway, the hp is wasted

oblique vale
crude tartan
#

ehhhhhhhh

dusty seal
#

What @crude tartan said. You need VRAM to do your models. The processing power of a 3060 will be quicker than CPU and DDR4 or 5 if you need to run models above 10gb.

drifting wadi
#

What about 2 1080 Tis?

crude tartan
#

10 series doesn’t have tensor cores so I think it’s a lot worse

gaunt glacier
#

vram isnt the only thing that makes a card good

#

have u considered clock speeds? RT cores? tensor cores? CUDA cores?

dusty seal
#

He’s referring to AI modelling. It has specific needs and priorities that out weigh other attributes of a GPU for the purpose.

#

Having 5090 level clock speeds, silicon and architecture would be functionally useless if the card hard 8gb VRAM for example. A 3060 12gb would be better for all modelling above 8gb VRAM.

#

In this theoretical scenario.

crude tartan
#

Because it’s THE limiting factor in how large a model you can run

#

Using any storage medium besides vram slows down performance a lot

warped flint
#

$500 is enough for a 4060 Ti 16Gb