>$300 brand new
>$250 second hand
Is the NVIDIA GeForce RTX 3060™ the best card for AI? How can anyone else compete
The King of AI Cards
>$300 brand new
Can it run Crysis on HD?
This is what I use for prooompting for now, waiting for a good deal on a Nvidiot card with at least 16GB of VRAM
why does nvidia torture us with such little ram. We need more, NOW. Like said, video games are for fucking children.
The only purpose of a gpu at this point is Ai.
This. I can run all the games I want with a ten year old graphics card, but AI's are what I need new hardware for. And for AI, VRAM is all that matters.
cos it would cannibalize their """high end""" offerings, which are literal gamer cards but with more ram.
you do realize 99% people still buy GPUs for games, do you? local AI is popular only with BOT coomers and schizos. 12GB VRAM is still plenty for gaming. you won't get more VRAM in consoomer cards any time soon.
G A M I N G G A M I N G G A M I N G G A M I N G G A M I N G G A M I N G G A M I N G G A M I N G G A M I N G G A M I N G G A M I N G G A M I N G
Nvidia agrees, but their actual AI cards cost $10,000 each and are being bought up as fast as they can make them. They simply have no incentive to try and go for the consumer market in any way at all, especially since the majority of gamers have no need whatsoever for any GPU more advanced than the RTX 3080.
Yes very well actually
The Intel Arc A770 has much better FP16 performance and the extra 4GB VRAM can take it further. Pretty much the only thing stopping it from becoming the value champion for AI right now is software support. Imagine what things will be like once oneAPI gets better support for PyTorch and Tensorflow.
VRAM is more important for AI. I can wait a few seconds longer for an iteration, I can't do anything about the memory.
> theoretical max
Now let's see the actual performance doing anything in the real world
Can run only 7B.
> the only thing stopping it from becoming the value champion for AI right now is software support
AMD shows it isn't that easy.
Not sure how they compare rn, but some time after Ryzen came out, Intel's software department had more employees than AMD had total employees.
Do these arc cards have dedicated AI accelerators like Tensor cores in the RTX cards?
Yes. They're called XMX cores (Xe Matrix Extensions).
Now do they support cuda?
go back to xeddit
Intel Arc is already killing it in Blender, they just merged initial hardware RT support while AMD still hasn't gotten HIP-RT for the RX6000 series out the door even though they promised it was going to be ready by late last year.
Have a 3060, bought if for unreal and it beat a 3080 in performance on that.
Now use it for prompting, but yeah I want MOAR vram. Looking at a 4090, prices are ridiculous here in EU
>Looking at a 4090, prices are ridiculous here in EU
They are already at or below exchange rate adjusted msrp + vat.
I'm waitgayging for one too but I don't expect them to go much lower.
Fuck off AI gay it's because of gays like you AMD limits the VRAM on other cards so AI gays don't just buy them all we'd have a 16GB 4070s and 20GB 2080s if it wasn't for AI gays immediately just buying them all to make porn all the 3060s being sold are all to AI gays wanting to make porn fuck every single one of you AI gays
>we'd have a 16GB 4070s and 20GB 2080s if it wasn't for AI gays
Eh, not like Nvidia started skimping on VRAM only recently - see 970 3.5GB. Also, models like 2080 TI and 3080 were popular for AI despite not having like 20GB.
>so AI gays don't just buy them all
Hey, I've been telling AI bros to buy Arc A770 16GB. As
says, for AI in particular it's got a very compelling price : performance (and VRAM).
You can do some gaming on it, too, and it'll age much better than the various 8GB Ampere VRAMlets.
(On a strategic note, I'd also much rather support Intel staying in the dGPU market than Nvidia's price policies. Vote with your wallet etc. Pic related.)
oh no no no no
recently 'upgraded' and went with this because of the 12GB. the next step was a 4070 for $250 more, and still 12GB.
>>$300 brand new
370 in germany
And euros are supposed to be stronger than dollars
19% VAT raping again
Remember, that's what the leftists want
yeah, buying electronics in EU and not having a business to deduct VAT is tough
vat is the most right wing tax, as it falls specially to the lower classes and doesn't consider the incomes or assets at all.
>he thinks leftists care about the average person
leftist theft is never right wing no matter how much you recoil, how injured you are, how you suddenly shrink back
you have to go back you retard
> this thing we advocated for turned out to not be what we thought it would even after evil right wingers warned us
> that means it's evil and right wing
You have a child's understanding of the world if you think "right wing" means needlessly attacking poor people.
it means my news channel wants me to like you and hate who they call "on the left" ooo spooky
I'm not telling
Literally just buy a second hand RTX A4000 with 20GB for €450
*king of AI cards for a hobbyist
there's your problem
t. creditmaxxed a used 3090 for 4-bit 30B miku
weird way to spell "based"
mr shekelberg just gives me stuff its awesome
so long as it's 0% and you can pay it off before it does get a rate it's always worth it
two assumptions that never hold true
don't spend money you don't have, or get hauled to court by the banks and watch them frisk you for everything you own
Spending money I don't yet have had been the best investment of my life
I bought my home
>dont spend money you dont have
good to know you are a financially illiterate boomer
you dont have to spend 3x your yearly wage in a month when taking out credit
money has a time factor
It is. I bought one used last December specifically for AI.
Hehe My Coom Generator
starving artists hate this one simple trick
Chinese vidya artists don't like it, either. Maybe except for the ones who get to stay at their jobs and use the AI to 5x their productivity or whatever.
>5x their productivity
still need to see that irl for any meaningful work
And the porn industry is in fucking permanent shambles
i don't fap or pay attention to 3dpd, is the mindgeek garden gnomes and friends actually worried
that's fucking nothing, although perhaps sufficient if "ai" for you is just diffusion cooms
>just diffusion cooms
I see you haven't been paying attention to the incredibly fast moving machine learning thread generals here
CPUs run useful LLMs brud
If you are looking for a GPU in the first place, it would be assumed that you are not satisfied with CPU speeds on textgen.
Can anyone explain VRAM to me?
>have 10GB card
>literally always maxed out
>even when I use the exact same settings I had on my 980 that had only 4GB of VRAM)
The fuck is this. I'm worried it's gonna keep filling up and then shift over to RAM which will cause my game to fuck up performance wise.
If it ran okay on 4GB, it's probs just hoarding assets in case it needs them again later, for better performance. Could also be reserving some VRAM for later use.
Because empty VRAM is wasted space. So they just fill it or mark it as full so nothing else can use it when it wants to. There is also a difference because "used" VRAM and VRAM actually in use.
VRAM capacity is only an issue when it has to constantly trash things in and out to keep up with the gameplay. The VRAM issue is mostly overblown by retards and a couple of new shitybox ports running like shit regardless.
Ignore any dogshit VRAM monitoring software, MSI afterburner for example can't differentiate between allocated and usage.
So that's why you can get two PCs, with the exact same in game settings but one has more VRAM, which have totally different VRAM measurements. Games typically just allocate all (or most) of the VRAM to use it when it needs to.
You'll know when your VRAM is legit maxed out because your game will stutter like a mess. But even if you see the VRAM shooting up, to near full or outright maxed, this doesn't mean the game is actually using all of it.
TL;DR - Ignore that shit, look out for stutters. No stutters = No issues.
>MSI afterburner for example can't differentiate between allocated and usage.
Funny, because RTSS (which is what MSI relies on for that overlay), can actually do that with the Memory Usage / Process checkbox.
RTSS is just the collection server. You still need a program to select what you want/display it. Such as RivaTuner or Afterburner. But it's been supported for a few years now.
see the second post replying to you
it's a good poorgay-tier card. the 3090 is king for mid-tier though
>$300 for ONE (1) component
That's basically the lowest you can go for a practical GPU.
3090 24GB (936 GB/s) == 10-20+ token/s (30B)
3060 12GB (360 GB/s) == 5-11 token/s max (13B) tuned
P40 24GB CPU (346 GB/s) == 4-7 t/s 30B | 11-14 t/s 13B (as swap)