Thread 106282598

>he has 16gb or les

we kek

>>106282598 (OP)
Nyet, 1050 Ti is fine.

Just three years ago 16GB was considered a lot.

>>106282651
and now it is not. what the fuck is your point

>>106282737
Maybe we're moving too fast.

>super
>18gb
doesn't sound very super. my 8 year old gpu has 11gb. 18gb sounds better suited for a 5060. 32gb for the 5070 super and 48gb for a 5080 super.

>>106282598 (OP)
let me guess $3000

>>106282598 (OP)
I hope you use them well haha

I really wish AMD didn't give up on high-end GPUs.

>>106285299
High end GPU should cap at the MSRP of GTX 1080 Ti.
Inflation is no excuse. My wage hasn't been inflated yet

>>106282598 (OP)
I still use a GeForce GTX 1060 6GB.

back in my day the mid range gpu option cost $250

>>106282846
No, the problem is 16GB isn't enough.
It should be.
That is the point.
The point is, 16GB isn't enough, while it should be.
Nobody is saying more shouldn't be an option. Or that more is bad. Or that progress should halt.
The problem is, 16GB isn't enough, when it fucking should be.

The people to blame are DEI hires and devs and publishers who refuse to properly optimize their games. Instead of optimizing and releasing a good product, they shove day 1 DLC down your throats and target 30 FPS with DLSS and frame gen on the latest card while maximizing the VRAM to deliver that shit FPS.

At this rate people will be renting entire streets to house datacenters and processing plants just to run a AAA slop title at 30 FPS in 2035. There won't be enough real estate on earth to house all the hardware needed to compute through tranny vibe code. With some optimization however we can fix everything but that's not profitable. More graphics cards sold, more DLC, more agenda. That's all that's on the menu.

Stop buying shit games.
Support indie devs.
Even 4GB is plenty on good games, and that includes non-upscaled 4K graphics with MSAA and texture filtering.

>>106285488
Chek'd as I should be.

>>106285353
Because it isn't out yet.

>>106282598 (OP)
Can't wait to get my 5070 Ti Super 24GB
>>106285299
Has nothing to do with high end, they just don't have CUDA and therefore will be inferior for most AI purposes

>>106285299
UDNA will fix it

reminder the 7900 XTX is a sleeper AI GPU and excellent value and VRAM for the price, as long as you use Linux.

I've been hodling since my gtx 280

is it finally time to upgrade?

>>106286046
>gtx 280
It's really been working for 17 years?

>>106286065
works great, all I play is old emulators of retro consoles

>>106286046
get a cheap card with 24GB VRAM. then wait for cheap dedicated consumer NPUS with >=128GB VRAM.

>>106286103
>cheap card with 24GB VRAM
>wait for cheap dedicated consumer NPUS with >=128GB VRAM.
I don't think that anon wants to wait 10 years

>>106285488
>he still thinks about video games
Chuddie, these things haven’t been for video games for years now, piss off. For real uses it barely cuts it

>>106282598 (OP)
192-bit bus on fucking 70 class, pathetic

>>106285488
(((Who))) hires the DEI devs and push unfinished slop though?

>>106286138
I think we'll have NPUs or GPUs with 128GB VRAM within the next two hardware generations.

>>106286233
I was focusing on the "cheap" part

>>106286158
Same as the 9070

>>106282598 (OP)
Those specs are pure fiction. GB203 only has eight memory controllers (256bit) nor can they clamshell it. Nvidia will not equipment their gayming SKUs with 3GiB GDDR7 chips.
Any future gayming SKUs that Nvidia decides to release with more 16GiB will have to be a highly binned GB202 chip.

>>106285488
16GiB is plenty of space for the vast majority of gayming settings and usage patterns until the successor of XSX/PS5 drops.
All of the "MAH VRAM" crap from /g/ are from poor-fag AI/ML types who can't afford the intended SKUs for it.

>>106286158
That has been a thing for while now for Nvidia ecosystem.

>>106286170
NJUDEA, of course. They're been grooming every gaming studio for decades now and always trying to get the industry to normalize whatever FOMO proprietary slop tech they're pushing at the time. Gaytracing, PhysX, whatever. Always an NJU behind it.

>>106286301
NOOO YOU'RE SUPPOSED TO HECKIN UPSCALE BRO YOU'RE USING THE WRONG SETTINGS
People like you should swallow a cactus.

>>106282598 (OP)
she, I am a woman, bigot.

>>106286384
>Brainlet /v/tard who doesn't understand how VRAM in managed in gaymes and simply sees big numbers in charts and panics it grows 10GiB+ for no reason

>>106286412
NOOO YOU MUST NOT PLAY WITH THE FULL RES TEXTURES UNLESS YOU UPSCALE!!!! YOU'RE ASKING TOO MUCH SAAR!

>>106286422
>can't get into LOD
>Insists on running content on utlra/epic settings that are little 90% flex-tier with very marginal gains in fidelity. Being forced to cut back hurts the ego far more than it impacts visual fidelity
100% v/tard, consoomer brainlet confirmed.

>>106286301
>poor-fag AI/ML types who can't afford the intended SKUs for it.
Yeah most people can't afford to buy enterprise grade GPUs. Have you seen the cost of a H100? It's like $30000+ for one. There's a reason they don't sell these to consumers since there's virtually no point. If you're serious about local LLMs it's far more smart to just offload to the CPU and accept the diminished performance. Having a high VRAM card still makes a big deal though. It's why most LLM fags buy old, high VRAM cap, cards and don't care about using the latest and greatest.

>>106282598 (OP)
I just want a decent 75w low profile gpu to run basic blender.

What are you people even using 18GB of VRAM for?

>>106282598 (OP)
>he doesn't have 10000+ cuda cores

>>106287362
PORN

>>106287362
future proofing for unreal engine 5 sloppa

>>106282598 (OP)
>18GB
What's the price?
>INB4 $899

Nooooooooooooooo I just bought a 5070 ti last month

>>106287536
get fucked lol

>>106287488
$549 supposedly

>>106282598 (OP)
We are awfully lucky that the only games worth playing are ones that don't require 8+GB of VRAM.

>>106287613
I will believe it when I see it

>>106287362
genning porn
can't censor shit when you already have it

Im using the 5060ti 16gb which is fine for all my needs and am using it to gen AI images and videos. I am glad I didnt go with less because it maxes out in the process.

>>106287628
You're poor

>>106286045
Its slower than a 3060, a 5060ti completely blows that shit out of the water in ai

>>106287710
>refurbished scalper price
retard

>>106282598 (OP)
Inb4 128 bit bus

>>106285536
yes it is
SMB can buy stacks of it right now

>>106287362
That leak is pure fiction and as well as "Super" rerfresh.

Nvidia has zero reason to release a Super refresh with Blackwell. At best, the RX 9070 is a 5070Ti competitor. Nvidia can easily do a price war if they were genuinely threaten by it (Protip: they are not, because of mindshare difference). Blackwell effectively replaced Ada Lovelace. All of the existing Ada Lovelace inventory is drying up so Blackwell SKUs will become cheaper to get.

>june 29
>no link
>>he has 16gb or less we kek

>anosn reply
>????

>>106288803
More like not stupid with money. You dense fool. The irony is that majority of those who get $1K+ SKUs are fiscally irresponsible poor-fags who get via financing or have to finance some other stuff. a.k.a deep in debt.
Actual Rich-fags ignore silly toys as gayming SKUs and are investing elsewhere to grow their lopsided gains.

>>106285299
High-end gayming SKUs are simply not sexy to investors. The market is mature and has no room for growth and is shrinking.
Nvidia uses it as a dumping ground for shitty yields from their professional line-up. AMD has a much smaller presence so they have even gave less of a reason to engage in it and rather allocate their production to more profitable Ryzen chiplets.

>>106282598 (OP)
I have an RTX 5080 FE. My plan is to simply resell it when the 6090 comes out and buy that. Then at that point, I should never need a new GPU again because we're never doing 8k.

>>106290481
Path tracing at 4K.

>>106290481
you underestimate how shitty games will run in a few years

They had 8 x 2GB vram chips on the cards before. Now they have access to 3GB chips. If every card is not updated to 1.5x memory, purely because of the new chips, then nvidia is just being extra cheap. They don't even have to change anything other than some programming on the card.

Fuck I hate nvidia. I bet they will still ship 12-16gb cards when they have access to 4GB vram chips. So shameless really.

>>106291372
Nvidia is protecting their ML/general computer line-up by intentionally gimping the capacity on their gayming SKUs. They don't want gayming SKUs being converted in "poor man's" compute/ML" units. That doesn't mean that will stop underground chop shops doing hardware mods to retrofit gayming SKUs with 3-4GiB GDDR chips.

>>106289901
The 50 series is like 5% better than the 40 supers because it was supposed to have this much vram in the first place.

>>106286045
Is rocm there yet?

>>106282598 (OP)
name 1 (one) reason to care about graphics in modern games? Name 1 (one) fun game that I can't enjoy on my 3060.

>>106282651
16GB is the bare minimum now. The ABSOLUTE bare minimum. Like that is what your $100 throwaway GPU should have.

128GB VRAM is what the high end would be if we didn't live in this stupid fucking duopoly market

>>106291525
Are you telling me that if I just go around grafting new VRAM onto old GPU's I can resell at a huge markup for big profit?

>>106291605
Now it's supposed to be a bad thing that there are yet no games that are completely unplayable on a 3060? lol

>>106291723
well, when such games appear and they are actually fun and not the millionth iteration of the same slop game studios have been releasing for years I'll think about upgrading, for now, no thanks.

>>106282598 (OP)
What's my 10th gen i7 2070 super iBuypower prebuilt worth if I were to sell on FB marketplace? Like a realistic amount.

>>106290688
Path tracing 4k is already "almost possible" on my 5080. Cyberpunk path tracing could be playable except DLSS4 performance mode introduces too many artifacts on textures with transparency with path tracing enabled. A 6090 would more than double performance, a 5090 is already like two times a 5080, 60 series should be new process and a bigger gen on gen gain than Blackwell.

>>106282598 (OP)
no thanks, I'll keep my physx gpu.

>>106291895
List it for $650 and accept no less than $450. Or chop it and sell the parts individually for more money and deal with the headache of using ebay

>>106291576
>rocm
DOA, why would anyone invest into this?

I enjoy my 3090.
Buy used 30 series cards.

>>106286275
You don't get 16 GBs out of 192-bit bus retard.

what bothers me so much is that since I do 3D work/rendering, I can't use anything other than nvidia. I would gladly use AMD or anything else as all I care about is vram but I'm handicapped to nvidia only due to rendering requirements. and alongside that, I'm fighting against all the gay ass AI niggers who tripled the price of cards the past couple of years. big fuck you to all of you

Give the 5070 ti a little more ram. Oh now we can sell it again under the super name!

>>106293032
Sure you can with the power of pcie 5.0, look at nvidias great 5060ti 16gb on 128 bit!

>>106293032
these two things are not related in the slightest
the core has a 192 bus to the memory units which then have their own wiring
in other words, you have an intermediary bw the core and the memory

>>106291677
>16GB is the bare minimum now
You are just telling me how much you really understand how video memory management works. 16GiB is plenty for the vast majority of gayming content and settings out there. The bar will not rise until successor of XSX/PS5 drops.

>>106282846
We're moving slower than ever. In the 90's your new pc was completely outdated 12 months later.

>>106286046
Impressive. For me it's an GTX 660 OC

>>106294766
Maybe two years after the successors drops. There won't be many next gen exclusive games at first.

>>106286400
YWNBAW GTFO TRANNY

>>106282630
2080 is better

>>106286400
>woman (male)
eww

>>106293000
>all the used 3090s are from mining rigs
wew, lad

>>106296633
Mining rig GPU's get stressed less than GPU's stuffed in some gaymer oven.
>b-but muh paste
Put PTM7950 and fresh pads on.
>b-but muh fans
Mine were fine, are easy to swap and so I did.
>b-but muh vram aboos
Mine is fine, I run AI workloads and it's fine.

By buying into the mining rig FUD you're cucking yourself out of affordable GPU's that just werk.

>>106285488
True. We also need stacked memory.

>>106287362
future proofing ai hallucinated code in games

>>106282598 (OP)
>flexing with extra 2GB
fighting for scraps I see

>>106282598 (OP)
Perfectly content with my 8GB RX580. More VRAM would be nice but I have priorities

>>106282598 (OP)
>270tdp
doubt.mp3

>>106282598 (OP)
>5060 gets 12gb instead of 8gb
Whoa that’s actually neat
>12gb 1440p cards like 5070 get 18gb
>16gb 4k cards like 5070ti and 5080 get 24gb
its completely pointless until like 2029 and by then they’ll be old trash

>>106285538
>Has nothing to do with high end, they just don't have CUDA and therefore will be inferior for most AI purposes
ZLUDA exists, or you can just build a HIP application that's multi-vendor and will work across AMD and NVIDIA.
>>106290028
AMD needs to work on merging CDNA with RDNA already. I know they were planning it. Then there will be no "High-end gaming SKU" but rather "Not good enough for a workstation/server card but still "high end" SKU"

>>106291707
22gb 2080 tis are nearly the price of what I bought my 3090 for (900 aud vs 1000 aud). 48gb 4090s are 5000 aud, just about the same as a 5090 (5500 aud).

It's amazing how everyone everywhere is just full of a mass amount of idiots.

The VRAM problem is two things. AI production supply/demand is eating up 5x as many cards as all the gamers combined. Also, it is being gatekept so people have no other choice but spending thousands per GPU to have good AI capable hardware.
If I could go out and buy $500 32GB 5060Tis, I would have four of them in my system since two months ago.
Instead the only choice most people have is $3000 5090s, or stepping up to prostation/server class cards to have 40 or 48 or 80GB per gpu, at a much much higher cost.

Also to the idiot who says you cant have 18GB on a 192 bus. 192 / 32 = 6, 6 x 3GB = ????
The 24Gbit gddr chips is old news, yes they are coming, yes they will be on Nvidia cards. Though I sincerely doubt they'll be on a 50 series ultramegasuperhyper Uranium-Meganium rerelease, but I'm wrong sometimes so who knows.

>>106297942
My thoughts exactly. Come back with 24 GB and I'll be interested.

Imagine being amd and seeing nvidia make infinite profit while your pile of shit enterprise gpus barely make a few dollarydoos because you were too stupid to spent a few millys a year over the last decade on codeniggers to unfuck your drivers
i wish no end of pain upon this company for all the annoyances their shitass software caused me throughout the years

>>106299008
Also people who talk about doing AI local on 16GBs that shit is a joke. 128GB barely gets you in the realm of "this is good at [extremely well defined task]"

>>106299022
The global talent pool to unfuck gpu drivers doesn't exist for multiple companies to be at parity. And even Nvidia famously constantly releases dogshit drivers.

>>106299022
They're probably content with all of the dough they're making from selling too many Epyc CPUs. Meanwhile NVIDIA's ARM efforts have still yet to bare fruits.

>>106299038
Nvidia ARM CPU will be with us SOON

>>106299073
But will it sell and more importantly will it be a platform for everyone or just some niche custom server thing (they already sell switches and servers today)

>>106299038
Yes I’m sure they’re happy they missed out on the biggest gold rush in tech in the history of the world cause they make some pocket money on EPYCs and desktop cpus

>>106299084
You're delusional if you think AMD could have possibly overtaken NVIDIA in any scenario in the past five years. They were on the brink of bankruptcy and irrelevance until Ryzen and Epyc saved the company.

>>106299022
>>106299030
how difficult would it be to release standards for development that prevent overbloated programs from making their hardware look bad?

isn't software the more obvious frontier for performance improvements at this point?

>>106299030
128gb gets you pretty far actually, glm air is really solid as a general model. All it needs is better tool calling desu
the only real exception is codeslop, for which you want paypig models

>>106299082
I mean, ARM doesn't run anything but basic stuff does it. Unless Windows on ARM emulation has reached new heights, I think it'll just be for some niche purpose.

>>106299096
>isn't software the more obvious frontier for performance improvements at this point?
Yeah, but devs are universally lazy and incompetent, so there's nothing we can do about it. The history of software past the 90s is the history of people being shit at code.

>>106299126
>ARM doesn't run anything but basic stuff does it.
That's more of an issue with software not targeting it rather than an issue with the architecture itself. If they come out with a good CPU and it gains critical mass then obviously software developers (both professional and people like game developers) will target it. But only if it gains adoption.

>>106299096
Is your argument “just let open source fags write the drivers for us”?
That doesn’t really work for enterprise which has maybe 200 companies worldwide as your target audience, and who have no incentive to release any of their software to anybody else. Joe Blow from Berlin isn’t going to hack on your MI gpu in his free time because Joe Blow has a $3k mac and not a desktop with $80k gpu

>>106282598 (OP)
If the 5080 super with 24gigs still gets outclassed by the 4090 I'm gonna force my nut in leather jacket man.

>>106299126
As of next year the entire apple lineup will be running on ARM chips, so to say ARM isn’t viable is silly

>>106285488
>The people to blame are DEI hire-
Stopped reading there
What the sweet fuck does this have to do anything?
Fucking hell /pol/ is a virus killing this website

>>106299157
i wasn't even aware open source tried to do that.

and i'm not just talking about drivers, but software overall. it just seems stupid that trillion dollar companies can't put together good programmers to best utilize their services/hardware

>>106299188
Good programmers are in short supply. Anyone who is any good will be a hiring target for someone. AMD's open source drivers for Linux are actually very good and they do amazing work but that's no good for enterprise, they need an AMD packaged solution with support guarantees even though that driver is often shit except for specialised workloads they happen to care about.

>>106299188
Everybody tries to squeeze out every single cent they can and software has been a cost center rather than a profit center. By the time amd realized they’re missing out on infinite ai money there was too big of a gap to overcome quickly
“Hey our programmers eat like 20% of our profit surely we can do it cheaper” is an easy thing to sell when you’re aiming for a promotion or to lateral to a different company for more cash, and the average manager cares about his own wealth infinitely more than whether RandomCorp prospers or not in the long term

>>106291763
The 3060 was a solid card but I got tired of not being able to play titles in my library because I run 4k. Sure I could play rdr2 or bannerlord. But I couldn't play kingdom come deliverance because it ran at like 25fps. And even american truck simulator was running at like 55fps. So I got a 9070xt. I'll upgrade it again in like 2 years.

>>106299022
They've killed intel at this point. It's getting a federal bailout. At this point they need to worry more about x86 dying and competing with arm more than competing with cuda in ai.

>spend almost $1k on 4070
>get 12GB
I really got fucked over on this one.

>>106299168
Apple is vertically integrated.

>>106299141
It will, especially since intel is dying. What should be frightening is aws graviton enterprise arm cpus. Ec2 runs on gravitons. Enterprise will start making custom arm cpus for specific workloads more and more and just cut out x86 entirely for more efficiency because at scale that makes sense.

>>106299093
This, they acquired ati in 2006 and were dying by 2014. I'm honestly surprised their gpus are as good as they are today. The 9070xt is the best price to performance available this year, and will remain so at least until price cuts and the 50 supers come out.

>>106299370
You should have gotten the microcenter warranty bro. Then you could have killed the card and gotten a 5070ti or 9070xt. I try to tell you guys, buy the replacement plan and then short the card out before 24 months for your money back.

>>106283449
>you'd probably have somewhere between 32-64GB of vram if it wasn't for the AI slop bubble screwing everyone

>>106299370
>spend almost $1k on 4070
fuckin how you dumb wojaknigger? I bought on release for like 600

>>106299422
>he doesn't remember the 970 3.5gb
Nvidia has a long history of giving you the absolute MINIMUM vram. Remember the 3080 10gb? What about the 1060 3gb? They did this long before the ai bubble.

>>106299360
Sure but theres fuck all money in cpu and theres truckloads money in gpu

>>106299448
welp, those were the prices around here back then. it includes 23% VAT - gotta feed the immigrants.

>>106299422
No you wouldn’t? Why would either nvidia or amd add vram which is completely pointless past like 16gb? It just adds cost and energy demands for practically zero upside once you have ”enough” ram

>>106299453
Cuda has no actual patent expiration date and everybody writes ai software for cuda not rocm. You have write for hip to run both. As an amd owner you have find forks to even run stable diffusion on amd cards let alone enterprise (by the way the ones that work on rdna 3 may or may not work on rdna4). It would be like if the railway already exists and it were illegal for you to make a train that runs on it because **le intellectual property** so you were asking everyone to rip it out and build new rails so your train could run. There is no way nvidia would make the same mistake intel did with x86 and let somebody else use those tracks.

>>106299012
might finna get 5080 super from my 3090

>>106299479
For me to render video of huge breasted futas at home

>>106282598 (OP)
I am still fine with my 8GB for now. My 2080 Super is still going strong.

>>106299008
Nvidia will not allow their partners to fit their gayming SKUs with 3-4GiB chips. AMD will likely follow-up the same suite for the reason.
To intentionally segment the markets and force those who have an use case for more VRAM capacity to pay the professional tax.
You see the same crap on the CPU front if you want more I/O connectivity, memory capacity/bandwidth then what normal desktop provide.

>>106289932
You're coping and mega poor, get the fuck out of the technology board

>>106299413
Nvidia has zero reasons to do a Super fresh with Blackwell. They got the market cornered and AMD has no real answer to most of Nvidia's gayming SKUs. Mindshare is hell of a thing.

>>106299449
970 is a 4GiB platform. It just had a partitioned memory space due to how it was binned. Nvidia only got heat for it because their marketing omitted this fact. Despite it later exposed by independent testing. 970 came out as a top-seller, as it had insane value.

>>106300242
>poor-fag projecting this hard and thinks silly ovepriced things like gayming SKUs are "status symbols"
Nvidia and AMD shareholders love to count on braindead paypigs to finance/pay for their shitty yields.

>>106287362
I'm generating my own degenerate furry porn i'm to embarrassed to commission anyone else with, especially for video generation.
https://files.catbox.moe/kdc6de.mp4

I wish vram prices in gpus were more in line with how much vram physically costs instead of being arbitrarily limited by nvidia who wants to suck money out of the consumers.

like 100gb vram gpu could cost <1k € if you just physically connected vram to a gpu chip

>>106286400
Go fuck another fly

>>106288809
I think both of those cards would OOM immediately with the wan txt2img workflows I am using on my 7900XTX

>>106300434
Im curious, what stuff did use to make this?
Never used SD stuff

>>106301032
Image base generated with illustrious nsfw furry pred + loras, then animated with wan2.2 high/low noise 14b Light verison. You can do it with lower vram but i would recommend 16gb+

>>106286138
More like 20

>>106282598 (OP)
>he has more than 6GB
lol.

>>106285488
hold up we have the CEO of optimization over here

>>106282598 (OP)
>18GB
kek Is this gonna be another 3.5GB situation where the last 2GB tank performance?

>>106300259
And the 50 series was supposed to use 3gb modules at release but samsung shit the bed and couldn't produce at volume. So you saw 3gb modules only on the laptop 5090s and their enterprise cards.

>>106301425
There are 6 memory nodes on the 5070 pcb.
2gb x 6 = 12gb
3gb x 6 = 18gb

Samsung is now producing enough gddr7 3gb modules for nvidia to do this. At launch they couldn't so they used 2gb.

On the 5070ti/5080 its 8 vram nodes on the pcb.
2gb x 8 = 16gb
3gb x 8 = 24gb

>>106300245
Other than the 9060xt being better value than a 5060ti 16gb and the 9070xt being the best price to performance card of this generation. The 5070 is only like 5% better than the 4070s and is not an upgrade. The 5060s have a narrow bus that makes them shit on pcie 3.0 which is who nvidia should be marketing to- they should be trying to get the people with 3060/3060tis to upgrade to 5060s but it just isn't a meaningful upgrade especially on pcie 3.0 where the lanes fucking matter. I don't know what they were fucking thinking this generation other than we didn't get the cards we were supposed to get at launch due to samsung.

>>106301425
No, they're just going to replace the currently in use 2GB memory chips with 3GB chips and as such get 50% higher memory capacity on the exact same memory bus configuration. It's nothing too weird or special, just higher capacity chips.

>>106285538
9070XT with 32GB just came out

>>106299022
Open sourcing was the first step, they just need to go further, but like you said, they're too cheap to take the risk and to foster the appropriate information and control to open source community to do it for them.

>>106285299
They gave up because they failed.

It doesn't really matter though, I can simply sell my previous gpu and buy a new one. Maybe I'll just keep the old one and leave it in the box. IDC, I'll just buy the new model LOL

>>106288803
What's wrong with Parasite Eve 2?

>>106302736
>Copium
Wrong, Nvidia intentionally did not want their partners to use 3GiB chips. Why can you /v/tards not grok this?
We will have to wait until Blackwell's successor for Nvidia to even consider equipping their gayming SKUs with 3-4GiB GDDR chips.

>>106303390
those same partnners that will all use 3gb chips on supers? k

>>106302933
Mindshare kiddo. Nvidia is too strong, Blackwell is only an upgrade option for users who are pre-Ampere hardware that didn't bite the Ada Lovelace bullet.
>>106303038
Wrong, it doesn't exist. The actual card that leaker mistook is a Radeon AI PRO R9700. AMD also has to protect their ML and professional line-up so they have no reason to equip their gayming SKUs with 3-4GiB GDDR7 chips.

>>106282598 (OP)
>16gb
> lowering the bar so OP's new toy meets the criteria

>>106303420
Only for their dedicated AI/ML SKUs. Not gayming SKUs. The amount of copium from Nvidiots /v/tards never ceases to amazes me.
This is no different then Intel forcing its desktop platforms to max out at 4-core SKUs throughout Lynnfield-Skylake era. If you wanted anything more you were force to bite the professional tax.

>>106303468
super cards will all use 3gb modules from all vendors

>>106303432
>Radeon AI PRO R9700
literally 9070XT die with 32GB moron, people already bought it, like diy users
https://www.tomshardware.com/pc-components/gpus/amds-elusive-radeon-ai-pro-r9700-makes-its-first-retail-appearance-for-the-diy-market-customer-on-reddit-buys-the-gigabyte-ai-top-variant-for-usd1-324
navi48 die moron
https://www.techpowerup.com/gpu-specs/amd-navi-48.g1071
same specs as 9070XT. eat shit

>>106303509
so it's not a 9070 then

>>106303493
There will be no Super refresh. All of the gayming Blackwell that you see what are Nvidia is going to be selling until its successor comes around. Nvidia doesn't give a crap about gayming market.
That copium is very potent stuff.

>>106303518
are you retarded? I said two times it's 9070XT die, what's wrong with you? it's a 32GB 9070XT, it's the same fucking GPU die, look at the fucking specs, 1:1 9070XT with double memory

>>106303531
>>There will be no Super refresh
lol this is why nobody takes you seriously

Last I checked Samsung memory is worse than Hynix memory for overclocking. For both CPUs and GPUs. Like it's no contest that if you game or do AI you buy Hynix products because they are faster. Samsung is for cope buyers like Nintendo Switch 2.

>>106303509
Yes, it costs 50-80% more then RX 9070XT you branlet and it also doesn't boost nearly as high because it is binned for power efficiency not extreme gayming performance. It is the same story for AI/ML SKUs on Nvidia camp that use the same silicon as their gayming counterparts but are equipped with more VRAM but cost 3-10x as much as their gayming version.
R9700 comparably low price is just AMD trying to do a price war at Nvidia's lesser AI/ML SKUs. It might work and some of AI/ML

>>106282598 (OP)
I just chill with my old integrated shit with 2GB VRAM, the old games I play occasionally don't need any more.

>>106303590
>overclocking
is that even a thing anymore?

>>106303548
well if the 9070XT 32GB is out then radeon buyers should be happy it's what they wanted and said would happen and so now then can all buy one to replace their 7900XTX

that's gonna happen

right?

>>106303590
In the real world it will make no difference when you're going to have extra overhead from all 4 channels worth of RAM in your system (maybe if you're gaming you'd use 2 sticks dual-channel only but not for general productivity).
For a single GPU (not a cluster) it may matter more though but it's unlikely to make a meaningful difference.

>>106303626
is 9070XT faster than a 7900XTX?

>>106303677
lol fuck no
it has marginally better ray tracing performance and a better upscaler because AMD software locked RDNA3 out of FSR4. Fun fact, FSR4 is also coming to PS5 Pro but never to RDNA3.

>>106303617
>>106303647
The difference between stock and tuned is massive and you're a fool if you say binning or ICs for dram don't matter.

>>106303697
FSR4 already works out of the box on Linux it's just a Proton command. It works with any game that supports DLSS through Optiscaler as well. It's only 3ms slower than Xess2 but that's still 10+ms faster than native.

>>106303697
>but never to RDNA3.
FSR 4 is already available on RDNA3, but not officially. many are using it already

>>106303716
not sure what you're high on anon, people undervolt these days and that's it. there's no gain from memory overclocking are you crazy?

>>106303737
You're a fool then. No point convincing a fool because his grand performance is already on display here.

>>106303716
I wouldn't call a 30 FPS gain (going from 5200 spec, you're probably using something like 6400 though, this is a best-worst-case scenario) massive. That's marginal at best, and you're by no means guaranteed that in every title (some games are more memory sensitive than others).

>>106303744
if I try to up my memory clocks I get shit in return and at some point the computer freezes and I have to restart it. people don't do that these days there's no real gain from it. just undervolt to get more perofrmance

>>106282846
nigger we should've been on 1TB VRAM cards by now and 256GB L3 Cache on CPUs if it werent for greedy stagnation fags

>>106282651
It still is, if you don't use GayEye or play UE5 slop.

>>106304187
Do we even have the technology to put that much storage space in the cache of a CPU/GPU yet and still meet the needed performance characteristics?

>>106304430
not with the current work force and management let me put it this way VRAM is dirt cheap yet nshitia purposely won't put any to hike prices on all their cards and ayymd is following closely behind with the milking

>>106304474
L3 cache is not just VRAM though it's VRAM on steroids and it's got to fit into a tiny package on the CPU/GPU.

>>106304569
CPU cache end goal is to replace RAM entirely in the future so it'll be possible but not yet.

>>106304398
it is a lot but these new GPUs are targeting the AI bullshit market

>>106304593
I don't think that will ever happen though because people will always have larger and larger datasets and tiered storage with different speeds will still be advantageous.

Can't fit it all in L3 cache? Then at least you can fit it in RAM even if the latency is higher. Can't fit it all in RAM then at least you can store it on a fast high capacity NVMe drive before batching to RAM.

There are real challenges for improving the speed/capacity of each layer of the stack though. I don't think it's as simple as saying we don't have these things because of stagnation.

>>106304647
>Can't fit it all in L3 cache? Then at least you can fit it in RAM
Yes that's how it'll go. Cache will take center stage and RAM will still be there but as an optional extension to it. Stagnation is also part the reason why we don't have new esoteric tech like I remember 15 years ago how all those new types of memory would hit the market and replace RAM and so on.

>>106304778
If anything that stuff from 15 years ago is becoming more relevant. Stuff like CXL:
https://en.wikipedia.org/wiki/Compute_Express_Link

>>106286153
shut the fuck up, poojeet

>>106282651
Nah it wasn't. It was only because for so many years we were stuck on 8gb vram. First 8gb vram card was releasedi n 2014 the R9 290X 8gb. And that cost 430 dollars back then. And we were stuck on that amount of vram until basically 2023 when it crossgen with the ps4 ended and all games were made for ps5. Then it was clear 8gb was too little.

If the industry hadn't been stagnant on 8gb vram for almost 10 years, the 30 class cards would have 16gb vram right now. While even a 5050 would have 24gb at least..

>>106282846
We are not moving fast. Issue is that for the past 10 years we have been too slow. In 2014 AMD released the R9 290X 8gb for 430 dollars. First consumer level 8gb vram card. 7 years later the RTX 3070ti was released and it also only had 8gb vram. Even though it was a 600 dollar GPU released 7 years later. A GPU that was 170 dollars more expensive than the first 8gb vram GPU still only had 8gb vram even with 7 years of technology advancement.

It wasn't until the 40 series that 8gb cards were relegated to the mid range offerings with the 4060ti being highest end 8gb vram card. And in 2025 Nvidia still released the 5060ti 8gb, full 11 years after the first 8gb vram card was released they still released a 60ti class card that only has 8gb vram.

>>106286045
My main concern is the Blender performance.

>>106301029
Don't follow AMD, just looked it up on amazon. $1000+? Might as well get two 5060 TIs for 32GB VRAM at like $450 each

>>106287362
I'll get the 24 GiB 5080 super or 5070 ti super, depends on the price.
My 11 GiB GTX 1080 ti served me well for 8 great years, time for a good upgrade and wait another 8 years.

>>106286384
>Gaytracing, PhysX
It's DLSS, TAA and fake frames now, get with the times grandpa.

>>106282598 (OP)
It's because they are using new 3GB chips. So it would be either 18 or 15

>>106286045
For me it's TUNIGU 306TT