site banner

Small-Scale Question Sunday for March 5, 2023

Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?

This is your opportunity to ask questions. No question too simple or too silly.

Culture war topics are accepted, and proposals for a better intro post are appreciated.

3
Jump in the discussion.

No email address required.

Currently on a 1060 6GB.

Buy a 3080. If you don't have enough money, save; anything slower is an objectively terrible deal. The 4070 will likely be a 3080 equivalent (just like the 3070 was a 2080Ti equivalent), but it's also going to cost as much as the used 3080s currently do anyway so it's up to you if you want to wait.

The problem with the 3070 is that it's 80% of the price for 60% of the performance; this is also true for the 3060 as compared to the 3070 and the 3050 as compared to the 3060. (This is the root of why tech bloggers complain about every new GPU launch, because it's not the "3080, but for 400USD" that everyone actually wants but can't buy.)

ML workloads function just fine on it; sure, the 4090 is twice as fast, but going from 1 iteration a second to 10 iterations a second is a much larger leap than going from 10 to 20 and it still can't run Facebook's LLM anyway so there's no real loss there.

I can't guess if nVidia has another doubling of performance in the pipeline for the 5000-series, but if they do it's probably going to cost roughly 1.5x what the 4000 series did because if they don't they'll likely end up cannibalizing their Tesla sales. Either way, it's probably not going to completely tank the resale value of that 3080, which will probably be a perfectly serviceable card far into the next console refresh cycle (which is the other half of the reason for why they're so expensive- the 8800GTX of 2006 was twice as expensive as the 7 series, but also had twice the performance and wouldn't be meaningfully topped for a very long time and as such was absolutely worth the initial asking price).

Also OP can consider AMD gpus if he wont use cuda. Depending on his market, he might be able to get a 6xxx card for cheaper than nvda counterpart, especially used.