Oh no, I’m sure no one could’ve guessed the names…
32GB is news. It confirms either a 512 or (more likely) 256 bit bus, which would be a significant drop from the 384 bit on the 4090.
I’m sure the increased perf of G7 would fully offset that, but this means without some larger caches it will be difficult for this to be a massive performance jump from last gen.
Unless they’re going 512 bit, in which case ignore all that and wow is this a monster.
More likely a 256 bit bus? Not with the monster of a card the 5090 would be. 256 bit on that massive of a card would be ridiculous. The card is set to draw 600 watts.
For the record, in case anyone cared, I audibly laughed when I read 600 watts.
sad PSU noises
Is 5090 the card model number or the price?
Uh it is both, for convenience, TLDR 5090
Doesn’t matter, I couldn’t afford a 5090 anyway.
going by trend, I don’t think my apartment could fit a 5090
You’d need a separate apartment just to keep it from sagging…
On the plus side you no longer need heating.
Nvidia doesn’t care. Their margins will be high enough, and there will be enough buyers that they’ll be able to print money at whatever absurd price point they choose.
It’s gonna be over $2,000. (Uneducated caveman ignorant guess)
If “Learning from history” counts as education, I wouldn’t call it an uneducated guess.
I’m seeing $1900, but guessing $2500+ once the scalpers get to scalping. I don’t intend to buy one, but I hope those assholes get fucked over.
Close It will actually be ~$8000
I got my 3080 and I plan on keeping using it for a good while still.
Knowing Nvidia’s exorbitant pricing, I think I’ll keep Intel’s Arc B580 in my wishlist.
They need a Steamdeck type list of games that will actually run on it though.
Allegedly, the 5090 would have 32GB and the 5080 16GB, I don’t see much room for the 5060 to have more than 8GB if the 5070 itself has 12GB?
I would have loved to see the 5080 with 24GB, the 5070 with 16GB and the 5060 with 12GB (at least). And for the 5060 to drop the 128 bit bus…
A 5060 with 8gb will have Intel wringing their hands with glee. Whatever performance gains the ($300-350) 5060 GPU has (and I don’t doubt they’ll be notable) will be choked off by an 8gb framebuffer. They might do a 16gb clamshell like they did with the 4060ti but like you said, where does that leave the 12gb 70 card?
I would have loved to see the 5080 with 24GB
They wouldn’t do this because it would undercut their 5090 sales.
Sales and probably pricing itself. With a gap this large, the price of the 5090 might be more than twice as high as that of the 5080 and people would still buy it.
If the value proposition continues, and based on the leak it seems like it will, the 5090 would be a better value FPS per dollar than the 5080 and anything below it. The 4090 cost like 40% more than the 4080, but gave 60% better performance. The 5090 looks like it will be well over twice the performance of the 5080, so I’d expect twice the price.
This is why I love Zotac. We’ll, not this reason, but it adds to the small pile of smiles they’ve given me.
That being said I’m skipping this Nvidia gen and might break for AMD next. My 3090 is still trucking fine and I feel like Nvidia has lost their value after the debacles of the 40 series.
skipping nvidia since gtx750ti, AMD all the way up
You should think about selling it TBH. 3090 prices are shooting up like crazy, and may be at a peak, because they are the last affordable card to self host LLMs.
Never even thought of that, is there a good website to sell a GPU on or is it pretty much just eBay?
I just don’t play games like I used to, just videos now. Poor thing hardly gets any use.
You could list it locally depending on where you are, through FB marketplace or Craigslist.
Otherwise, yeah, eBay.
Can’t you run LLMs on 4090/5090 maybe 5080? Basically any Nvidia card with 24GB+ of VRAM?
Yeah, but they not worth it.
The 4090 is basically just as good as the 3090 because it has the same amount of vram, but twice the price… so you mind as well get 2x 3090s.
The 5090 will be hilariously expensive, and 24GB -> 32GB is not that great, as you still can’t run 70B class models in that pool… again, mind as well get 2x 3090s. I would not even bother trading my single 3090 for 5090.
If AMD sold a 48GB consumer card, you would see them dominate the open source LLM space in a month, because every single backend dev would buy one and get their projects working on them. Same with Intel. VRAM is basically the only thing that matters, and 24GB is kinda pitiful at a 4090’s price.
AMD is by far the price-per-unit-performance leader. Same is true on the CPU side. Intel and Nvidia aren’t even on the chart tbh.
I just want one to self host a 70B LLM model for fuck’s sake. I don’t want to be forced to take out a god damned mortgage/personal loan to buy one.
I picked up a pair of old Tesla P40s. Right now I’m running a Q4 quant of Qwen 2.5 72B that fits in the combined 48GB of VRAM with 12k context. They aren’t as fast as newer consumer cards, but it generates as fast as I can read while costing less than a used 3080.
interesting. They are cooled passively, right? What’s your case and cooling setup?
Qwen 2.5 32B is where it’s at now. 24GB is affordable, and it fits perfectly.
Otherwise, stay on the lookout for AMD Strix Halo, which can reportedly allocate up to 96GB on its IGP, and you can run faster backends like vllm or exllama.
What’s up with Qwen that makes it better than anything else?
It’s just smarter with the same number of parameters. Try Qwen QwQ or Qwen coder 32B, see for yourself… it stacks up well against huge models like the 123B Mistral Large, or even GPT-4.
Why? Alibaba trained it well, presumably with better data than OpenAI or whomever else, though specifics are up for debate. Some suggests that bilingual training on English/Chinese (aka the two largest text corpuses in existance) significantly helps the model over mostly english. Some say the government just gave them better data. There’s also suggestions that having so few GPUs compared to American AI companies made the Chinese “thrifty,” and gave them far more incentive to be innovative rather than brute forcing models (which has diminishing returns).
Damn, they require like 50GB vram, that’s nuts.
You might just want to use Kaggle tbh
Never heard of it.
