Hacker News Reader

Comment by greatNespresso | Hacker News Reader

Is Nvidia RTX 4000 solid for training? Alternatively this card seems listed for around $1800 so at $180/month on Hetzner, it may be even cheaper to buy it, depending on the expected usage.

View on HN

No, not a great GPU for training. Consider something like A5000, A6000 or their Ada variants - if you want to go full "data-center", A40, L40, L40S are solid picks for small-scale training.

Why does this GPU cost so much? It seems to be like an RTX 3060 Ti/3070 with more VRAM and less power draw. Does that really justify the price tag?

NVIDIA's licensing. Note that the 3060/3070 is a "GeForce" whereas a 4000 is a "Quadro". Then look at the GeForce driver EULA: https://www.nvidia.com/content/DriverDownloads/licence.php?l...

> No Datacenter Deployment. The SOFTWARE is not licensed for datacenter deployment, except that blockchain processing in a datacenter is permitted.

I could never figure out why they allowed blockchain processing, given how wasteful much of it was and how it impacted card supply for years, but there you go.

Because they made a lot of money selling to miners at prices that maybe they couldn't ask to gamers?

Wouldn't the GeForce license then say "no block chain" and a different NVIDIA brand (BlockForce?) would allow it, letting NVDA charge more for large scale mining? Otherwise it looks like gamers and miners are paying the same prices.

They do. But since most miners would just ignore the license, they implemented LHR (low hashrate) beginning on GeForce 3000-series that detects the types of ops used for mining and artificially slows the GPU to about 25% of its actual speed.

They did that but with mixed success. One random link from a Google search: https://old.reddit.com/r/EtherMining/comments/t4mwxj/workaro...

Their source code got leaked and it was reverse engineered to defeat it. That said, it was all kind of late in the game, GPU mining died not too long afterwards.

I bet that now the AI demand spike will be the new crypto demand spike that NVDA latches onto.

This seems insane to me. It makes me even more confused that AMD/Intel aren't trying their hardest to go after AI compute. Nvidia is basically leaving the door open.

But I suppose that the idea could be to protect their gaming market in a way. If datacenters were buying up gaming GPUs then that could cause shortages for regular customers.

AMD is trying their hardest.

Because it's "enterprise".

Datacenter customers don't have options, gamers do.

RTX 4000 isn't even data-center grade - it's workstation/pro segment. This is an important difference, as some server vendors (such as Dell) will not validate the workstation segment to run in their servers.

The segmentation is a huge mess from NVIDIA. Consumer grade GPUs aren't supposed to be run in data-center / compute (forbidden by EULA) and the workstation GPUs are in a weird middleground limbo, being many times more expensive, but not exactly attractive for server use for the above reasons.

Like Apple, they use RAM for market segmentation.

Are you going to be training constantly?

TBH renting a big GPU (A100 or better) and getting the run over with quick is usually best. VRAM is everything.

It's with ample RAM and the latest Intel CPU. Assuming the whole thing draws 100W, that's also including 72kWh of electricity.

I'm not sure what the exact 'rule' is here, but you wouldn't use a contraction in this sentence.

It is, with ample RAM and the latest CPU.

From https://www.yourdictionary.com/articles/contractions-correct... -

Contractions can be used in any position in a sentence; however, homophone contractions such as "it's" and "they're" sound better when followed by another word or phrase. The reason is that the sounds of "its" and "it's" and "they're" and "their" are so similar that they can be confusing unless they are used with the context of an additional word.

I think it's less to do with homophone confusion and more to do with usage patterns relating the contracted part to a phrase rather than a single word [1] [2]. By the logic of the homophone explanation, the following exchange should be much more acceptable than a sentence ending with "it's" or "they're", but to me it seems equally strange:

"Are you waiting for someone?"

"Yes, I'm."

[1] https://www.youtube.com/watch?v=CkZyZFa5qO0

[2] https://en.wikipedia.org/wiki/Clitic

To be fair, the contracted version doesn't read nearly as nicely.

    Q. Is it good for training?
    A. It's with ample RAM and the latest Intel CPU.

    Q. Is it good for training?
    A. It is with ample RAM and the latest Intel CPU.

I think they are agreeing with it not reading as nicely. But you still need the comma. Otherwise it sounds like:

"It is with great pleasure..."

I don't think you necessarily need the comma but it would depend on the reader, I think. I didn't need the comma to read it correctly but I can see how you might mentally insert a pause between "it is" and "with" that makes it read weirdly.

Probably better to reorganise it as "With ample ram [...], it is."

You have to discount the price of a similar server without the GPU, tough.

But even so, it's still quite expensive. I have the impression Nvidia charges more from data center customers. So, for cards available to retail consumers, it's not a fair cost comparison.

With only 20GB, RTX 4000 has limited use... I was expecting they'd be offering something with at least 60GB at this point.

Reply on news.ycombinator.com