Blame memory manufacturers who haven't increased the module capacity in 7 years. GDDR6 was 2GB in 2018, 2025 and GDDR7 is still 2GB, go figure why they have to pull out a crazy 512 bit bus on the 5090 that they haven't used in 10 years.
What about AMD? Stuff like MI300 uses HBM as well.
https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html
> Dedicated Memory Size 192 GB
> Dedicated Memory Type HBM3
Workstation GPUs that use GDDR6 (consumer designs with special 'workstation' drivers) from both AMD (W7900, uses 7900XTX chip) and NVIDIA (A6000 Ada, uses 4090 chip) top out at 48GB. Both have 384 bit bus and run 24 2GB chips in clamshell mode -- so absolute maximum possible for their bus width with existing memory chips. Stuff above that amount of memory is designed for server from ground up and uses HBM.
Nvidia could easy do more mem it if they wanted. There's another side of the PCB to use. They're just taking all the gains from trade the past 6 years since they have no competition and skull f\*ck1ng consumers, plain and simple.
They could easily put more memory on the 5090 the same way they have been putting more memory on the V100 A100 etc for 10+ years now.
They've just hoarding value and refusing to bring innovation down from higher end products to consumer products at any point over the past 4 generations now.
TBF the 3090 had memory chips on the back that did cause some (minor) issues.
And the *100 class data centre stuff all uses HBM which isn't comparable from a packaging pov
But yes, they should probably have +8gb on the 80 and 90 class cards and +4-6gb on 70 and bellow for this generation
That's due to covid really. To develop denser dies, they have to build a new factory. That factory building infrastructure was going full steam until covid threw a wrench into it's gears and ground it down to a halt.
It'll take some time to build that momentum back up again. It's not like they're intentionally trying to not make money. The market demand is there and everyday they don't fill it they're losing potential revenue. It's not intentional. It's circumstantial.
Nvidia was literally doubling the memory capacity every generation when new memory came. 3GB 780Ti, 6GB 980Ti, 11GB 1080Ti, 24GB 3090. After 2018 it stopped with 2GB GDDR6 and 6X and since then the memory capacity hasn't increased FOR 7 YEARS. There is 0 reason for them not to use the new memory other than product segmentation, which is logical. Putting the same amount of vram on consumer cards as workstation ones would make render farms buy up 10000x of these which would be just a tiny problem.
In fact Nvidia is actually ahead in this direction than even AMD. Their cards already have GDDR7 controllers. When the new 3GB modules come it's just a question of replacing.
AMD has no way of upgrading their cards as it was confirmed they'll use the old GDDR6X.
it seems the gaming cards are being designed for gaming.
nvidia also announced project digits: a 3000$ pc with 128GB unified memory fully designed for models. two of them can run a 400B model.
it makes sense to buy that instead of a tiny 32GB card full of calculating power only for the VRAM.
People run their models on dedicated servers anyway for security reasons.
that fully makes sense to me
Yeah, I'll skip this series. My bottle neck is AIOPS it's memory. I'll just wait and see if the 4090 drops in price at all. Plus 36G is just a strange stopping point if they were really going after the prosumer AI home market. It's more than you really need for a fast 30B model, but not big enough for a 70B model even at a very low quant. Just odd.
lol you still can't even buy a 4090 from best buy. You might be willingly skipping this series, but you will be doing it regardless if you want to or not.
If you can't buy a 4090 from best buy right now, what makes you confident you will get the 5090 or the 6090 without paying scalped prices?
Im happy for you as do i myself, but i mean seriously go on best buy and let me know if you can add a 4090 to your cart.
The same fate will happen to the 4090, 5090, 6090. Scalpers have taken over and it will never be the same.
I think they intentionally make both the memory (16GB vs 32GB) and price (999 vs 1999) half of RTX 5090 so that people would just buy the 5090 for AI. Only need 24GB? Nope, sorry, buy the 5090.
GDDR7 VRAM must be more valuable than gold or printer ink. [8GB of GDDR6 VRAM is $25.](https://www.extremetech.com/gaming/gpu-vram-prices-drop-to-just-25-for-8gb)
> 8GB of GDDR6 VRAM is $25.
Those 1GB modules would allow for a 6GB 5070, 8GB 5080 and 16GB 5090.
The VRAM chips are cheap, designing a GPU core with a wider memory bus to accommodate more modules is expensive.
Yeah when the card sells for like $20,000 you can do a 80GB stack of HBM2, but that's not a viable option for a consumer GPU. You can also do backside VRAM in clamshell mode, but that cuts the per-module bandwidth in half and isn't ideal for a gaming card.
I definitely saw HBM memory in consumer card for pretty low price - Radeon VII. So HBM memory is also viable option. One and only reason to not produce consumer cards with 48GB of memory is monopoly.
Yep and you mentioned cards with a 5x larger stack of HBM, which sell for +20x more.
The VII was also a niche prosumer product with costs subsidized by Pro & Instinct cards with price tags far higher.
>One and only reason to not produce consumer cards with 48GB of memory is monopoly.
Nah, it's that mainstream consumers don't have any reason pay what it costs to design GPUs around that much VRAM
>Nah, it's that mainstream consumers don't have any reason pay what it costs to design GPUs around that much VRAM
Those who buy Apple do pay so there is a market.
Yeah 5090 is clearly an AI prosumer card, while all the new DLSS 4 or whatever features Jensen was hocking sound not VRAM intensive. They are trying real hard to push gaming towards lower VRAM so they can keep sales there high while raising the price potential for the AI hobbyist and small business niche.
Or maybe you know, Micron and Samsung can move their ass and make an actual progress with memory.
Ya'll here blaming Nvidia, but GDDR6 has had 2GB modules now for 7 years, since 2018. I'm not joking. GDDR7 is still just 2GB after 7 years and people still sit on "Nvidia greedy" while the situation is so bad they have to pull out 512 bit bus so their top end card can have more vram.
At the scale Nvidia buys at you tell the manufacturer how many you need like a year in advance and they make them for you. You're basically purchasing machine time to make the things, not ordering from a wholesaler.
Yes and there are practical limits to even to that. It's not a magic the manufacturer still needs people, machines, their suppliers etc. Especially with cutting ovládacích they technology they cannot necessarily make as many as they'd be able to sell
1 single 5080 SKU would have sufficed, as 2GB is fine for 16GB cards and below.
Or they could have not done the scummy thing with mobile GPUs; by first unifying them with Pascal (10 series) so they used the same die as the desktop versions, only to later go back to using lower tier dies for mobile GPUs while keeping the same names as desktop counterparts.
You might have replied to the wrong comment but anyway I agree completely. The RAM setup for these cards, especially 5080, look like they would be natural upgrades to Ada Lovelace if 3GB GDDR7 was widely available. It will likely have to wait for the Super refresh or maybe the Pro RTX cards will get it first to make for something like a nutso 96GB B6000.
My thoughts exactly. It's not like there's a bunch of memory chips that never get purchased by NVIDIA, AMD, etc. Those memory chips are all getting bought up and used up downstream.
It's not just about pushing gaming to lower VRAM. We're nearly 5 years into the current generation of consoles, and the next gen aren't hitting until 2027. Games are made for the PS5 first, they aren't built to need 16GB of VRAM.
5090 is an AI card. They know it, everyone else knows it. It's not for gamers, there's no market for a $2k gaming card when games are specced for consoles that came out nearly 5 years ago.
They should be more VRAM intensive considering they're moving to a transformer model, where the space requirements are usually heavier than CNNs. They're also doing speculative decoding, or some kind of multi frame generation which should intuitively have higher space requirements..
Bandwidth is the constraining factor for DLSS, not memory. DLSS is a tiny neural net - AMD's equivalent is something like 2MB. It has to be to run in 4 ms. But you have to feed millions of pixels into the net, which takes bandwidth.
Yeah, the 5090 is way overkill for a gamer card, no ifs-ands-or-buts about it.
That said, 16GB for a top tier gamer card is still a crime. With next-gen consoles likely to be 20+GB of addressable RAM for video processing, It's going to severely limit the longevity of that cards life.
Nvidia should have gone 20GB for the 5080. Low enough to dissuade AI hobbyists and make them get the 5090 (or a used 4090), high enough to put no doubt into the longevity of the card for gamers.
> the 5090 is way overkill for a gamer card
(x) to doubt
Some people out there have 4x the pixels to push at twice the framerate as whatever 60fps 1080p panel youre using
... wont need 32GiB of VRAM.
gaming will always want more processing, cant argue with that. But 32GiB of VRAM wont be fully utilized for 8 years minimum in the gaming sphere, probably not until the PS7 generation of consoles.
All I'm saying is a gamer would pay the exact same amount for 24GiB of VRAM with the 20,000+ cuda cores. Having 32G is a distinct nod that this is a prosumer card.
No need to be combative, for 5000 series we only have a choice of 32GB or 16GB, you alone started saying that 24GB is enough, but that's not a choice for 5000 series. I'm simply pointing out that 32GB is not overkill since 16GB is not enough, this is only about 5000 series, I am not talking about 4090, 3090 Ti or 3090.
The bad news is that I think they will go up and 4090s won't drop in price.
16GB on the only affordable 5xxx cards may even push some gamers back to the 3090.
I'm looking for a 3090 after the 5090 launch yet the hq still makes sense here hahahah (it will be the 1080ti destiny when I can get new-not-so-new hardware :D
fun fact: I only got the 1080ti back in 2017 bc I wanted to study machine learning/dp (but is not my main career so I couldn't dive deep in mathematics n stuff :(((( )
Lol, I got it for gaming back in college (took me 1 year to save money). But just wanted to see what all the ruckus was about AI and Nvidia ran some online Pytorch code for benchmarking (July 2018) and haven't been able to touch grass since. Lol
Should I tell my wife that I have no choice but to buy it for future AI research, or should I just be honest and beg her, saying I want to run old games at the highest settings?
Well, at least they learned their lesson with the 80 series prices… though they are definitely going to make an 80 super later this year for 13-1400 (and 20-24gb vram).
> (and 20-24gb vram).
It will be exactly 24gb. The next generation GDDR7 chips are already announced at 3Gb (vs 2Gb now) so all VRAM upgrades will be at 1.5x scale.
My point still stands. RTX 6000 Ada starts at roughly $8k, so making a 48GB 5090Ti will still undermine their profits from professional lineup, so this card won't happen.
what do you think of those [digits ](https://nvidianews.nvidia.com/news/nvidia-puts-grace-blackwell-on-every-desk-and-at-every-ai-developers-fingertips)units?:
"The GB10 Superchip enables Project DIGITS to deliver powerful performance using only a standard electrical outlet. Each Project DIGITS features 128GB of unified, coherent memory and up to 4TB of NVMe storage. With the supercomputer, developers can run up to 200-billion-parameter large language models to supercharge AI innovation. In addition, using NVIDIA ConnectX® networking, two Project DIGITS AI supercomputers can be linked to run up to 405-billion-parameter models."
1 peta flop at fp4 sounds a lot like 500 tops at fp8, so roughly 4070 levels of perf at 128gb vram for $3k. Less powerful than an m4 max, but also 66% of the cost (or 150% of the cost of a 64gb mac mini m4 pro)
LPDDR5X. Nvidia can do it if they want, but it costs money and engineering effort. They go up to 512 GB/s on their original 72 core Grace CPU which they use in a much more expensive product: https://resources.nvidia.com/en-us-data-center-overview/hpc-datasheet-grace-cpu-superchip
Seems interesting - a competitor to apples m4 lines, likely to try and claim some of that market.
But I don’t know about the speeds, prompt processing supposedly takes pretty long on these kind of systems (not entirely sure how long, would be interested if anyone has a good source on that), so I’m probably just going to get the 5090, and then stick another card in to get 48-56GB total vram.
I'd wait until you see that neural rendering tech in a game before you buy. That's a frontier technology and it's going to take a while before you see it deployed in-game (especially if it doesn't work on the consoles, which most games are developed for first).
all of this makes sense, not saying i like it but i see where nvidia is going with this. they have a war on 3 fronts: consumer gaming, AI inference/generation, and AI training. on the gaming side, they have to contend with intel, amd, and their previous gen cards. more gamers are realizing they're older cards are more than enough for 90-98% of games on the market at high/max settings, which might explain why intel's cards are only trying to match the 4070 in terms of raw gaming power. amd is poised to release another gen of cards, and the 4000 series is so good that honestly there may be tons of folks not wanting to upgrade (hell the 3000 series is still more than enough for most new features in games).
on war front 2: they have AI inference and generation. they see there is a fast growing market for local AI with average consumers and they see apple has the jump on that market with the their M-series of chips and unified ram. the one thing apple isn't doing is directly targeting those types of customers, but instead they are plugging apple intelligence. So nvidia is "striking while the iron is hot" and directly targeting those consumers with their digits project.
and on war front 3: they still have dominance with their A series cards among big tech firms and startups that want to cash in on the AI gold rush. this is a delicate area because not only do you have competition from AMD that wants to dethrone you in that area, but you also have consumer grad cards that are currently being used for inference and generation. they can't let these cards compete with their highend A series of cards that cost $20k+. I don't expect to see higher vram in their future gen gaming cards for local AI, just enough vram to run AI that's built into the card for DLSS and other systems. Even though this sucks, their project digits seems very promising, either that or an M4 mac mini pro. Either way, I feel less inclined to buy a 5090.
I'm really waiting for some tech reviewers to get a hold of the 5090 before launch. I feel like those reviews are going to make or break that card (and maybe the whole 5000 series). which doesn't really bother me. My 3090 is a beast and honestly ray-tracing in 4k wasn't enough to make me upgrade to the 4000 series and even though i like that they have a new card in a sexy slim 2/2.5 slot configuration, I'm on the fence about whether or not i even need the performance of the 5090 if digits is where they are trying to steer AI consumers to.
Time will tell i guess.....
Must be for a reason then, honestly AMD abandoning the high end to the point where they're not even outdoing their last gen top card just spells doom and gloom all over.. Nvidia can basically do whatever they want now
to people complaining about the low VRAM: it seems nvidia foresees the gaming cards are being designed for gaming.
They learned this from the time when cryptominers dryed out the gaming market for the GPUs.
backl then they reacted by bringing gaming GPUS that were hardware-castrated for mining.
now they are doing the same with AI to protect the gaming sector:
nvidia also announced project digits: a 3000$ pc with 128GB unified memory fully designed for models. two of them can run a 400B model.
it makes sense to buy that instead of a tiny 32GB card full of calculating power (that AI people dont need) only for the VRAM.
People run their models on dedicated servers anyway for security reasons.
that fully makes sense to me and i think this will bring a similar change where gaming GPUs are not going to be relevant in a near future.
it would be of course great if they would bring a 2000$ AI card with 128GB memory... but i guess the current PC architecture does not allow for that so it easier to design their own hardware (which is part of their core skills)
if all you want is gaming a 200$ card is enough.
the thing is that for gaming you need a fast GPU and some VRAM.
for AI you need fast VRAM and some GPU
you cant currently get both.
when gaming started people were trying to do all gaming with the CPU and soon realized that you need dedicated hardware.
Just like you people at first were complaining about buying a second GPU when all motherboards had an integrated gpu (which they still do!). thats when people started to buy gaming consoles for dedicated gaming.
the parallel is:
back then CPU mediocrely doing both tasks: today GPU mediocrely doing both tasks
back then dedicated new hardware for new task: gaming console: today dedicated hardware for new task AI PC
it seems current PC architecture does not allow for cheap unified RAM.
im sure in a near future motherboards will allow for it and we will get PCs with fast unified RAM.
hint: Apple architecture does allow for it. thats why M4 Studios can load and run decently large models.
I have a single PC, with 128 GB SUSTEM RAM, high-end i9 proc, plus 4090.
My motherboard will support only one card at x16.
One card at x8.
If I place the 5090 in the 16 slot, and say the even stands a chance I can fit the 4090 in at the same time;
1) If I putting both cards at risk, so close to one another, the heatness gonna meltdis?
2) Is the 4090 even worth running on at 8x channel when it needs 16x?
This is the Recommended retail price of the founders edition.
Expect retailers to mark it up to market value. Expand your budget because it won't stay within these margins. We going full bleed on this generation.
I don't get the complaining. $1999 for 32GB VRAM is far less than I was expecting. I was seeing $2000 4090s at Best Buy last summer. I though the consensus was the 5090 would be over $3000.
Unfortunately, my job is pretty easily replaceable by AI.
On the brighter side, maybe Llama 25 can buy itself a used 5090 in 2030 with the money it earns from my job :/
Shit, now I really don't know what I want. I have to choose between one of the following:
A second RTX 8000 Quadro for 96GB VRAM
A 5090 alongside my RTX 8000 for 80 GB VRAM.
If I went with the second option, that would make the RTX 8000 Quadro the display adapter, while the 5090 focuses on the AI workloads, but its common for me to exceed 40GB of VRAM but not 48GB VRAM, which would cause Ollama to cut into the RTX 8000 Quadro's VRAM after maxing out the 5090's VRAM capacity when running mutliple models simultaneously.
On the other hand, if I choose a second RTX 8000 Quadro, I would be getting 96GB VRAM and still run models at decent speeds with higher capacity, but I would not be able to take advantage of newer-generation inference upgrades, like flash attention 2, etc.
What would you guys do?
Sell the 8000 while you can get $2k for it and buy 2 5090s. Well assuming they stay in stock and actually cost $2k.
A 5090 is just so much faster than a RTX 8000.
Hmm...but that would lower my overall VRAM to 64GB. Its like choosing a bike for higher top speed vs acceleration. And one of those 5090s is going to have a noticeable slowdown as the display adapter if I were to be gaming and running AI models locally (which is something I do every day).
I'm predicting that used 4090s are about to drop in price by half. Used 3090s by 10%. Used 4060ti by 25% possibly making them the best budget option over 3090.
Let's see, prices for 4090 have gone up by 25% in the last few months due to Nvidia restricting supply. I think prices will hold as there is no competing 24GB model, either you pony up more for 32GB or drop down to 16GB. Jensen is diabolically clever.
Look... maybe unpopular opinion here but these cards are just totally shit.
32GB VRAM isn't enough memory to run anything interesting and leaves the cards essentially permanently crippled at these prices. The rest are even more useless.
Really the best thing about these cards is that they'll (hopefully) drive prices down for the 4xxx generation such that we can afford two or three with enough VRAM to be useful. No chance any L4's will hit the market at a reasonable price ever.
the problem is consumer and professional usage has such an astronomical divide in pricing. Consumers don't have the money but companies and research don't give a fuck whether the card is 2k or 10k. Nvidia obviously wants to either get people to buy many 5090 or their 30k h100.
They will prop this up until their hand is forced. (e.g sony closing the gap with consumer cameras and their professional cameras)
32GB can run just about every image gen model, and it’s enough for half-decent lora training on them.
In terms of raw inference time, the 5090 would be extremely fast - particularly in knowing that the professional series gpus will likely see their base clocks halved.
In terms of LLMs and multimodal models though, yeah, you’re not going to get far with 32gb. Super fast inference times, but you won’t be able to load much of a model in there.
I think it'll have the opposite effect on the 4xxx cards second hand market. A lot of people were waiting on a hope and a prayer that the 5xxx series would have decent vram.
They said the CUDA cores and AI Tensor Cores can now overlap computation! Also new version of DLSS
RTX 5090 $1,999 3,400 AI TOPS
RTX 5080 $999 1,800 AI TOPS
RTX 5070 Ti $749 1,400 AI TOPS
RTX 5070 $549 1,000 AI TOPS
I'm assuming AI TOPS is FP4 with sparsity? (So 1,700 FP8)
If RTX 4090 is $1,599 (1,321 FP8 TOPS) then RTX 5090 $1,999 is 25% more expensive - and FP8 is 3,400/2 = 1,700/1,321 = 29% more.
I guess the price is on the low end of most rumours ($2,500 some rumours)
$2500 was a nonsense rumor since pricing for this tier is determined not by performance but by the demand of enthusiasts which has proven to be below $2000 in spring 2022 dollars with the 3090 Ti flop.
They've been doing it for years with the data centre class GPUs, just look at any graph of performance for V100, A100, H100, H200 and you'll see on each generation they claim double the perform but at half the number of bits/float
Let’s look at 16 bits. 4090 was 330 and 5090 is 26% more at 419.
If you compare B100 to H100 the jump is 76%.
Basically the uplift is one third of the data center chip uplift. If they had competition, it would be 50-75% on the 5090 too.
Actual TOPS: 3352, 1801, 1406, 988
(from the official press release: https://www.globenewswire.com/news-release/2025/01/07/3005067/0/en/NVIDIA-Blackwell-GeForce-RTX-50-Series-Opens-New-World-of-AI-Computer-Graphics.html )
I sat there waiting for the reveal, knowing full well that it was going to be ridiculous, and then I was still blown away. An increase of 400usd in a single generation is insanity. How can it be twice the price of the 80 series?
That is not true, there is a huge upper middle class hpc/gaming userbase that are in high earning jobs willing to fork over this money.
You need to understand that when it comes to hobbies, computers are on the cheap side. Racing, boating, flying, golf etc can easily cost way more, especially since they are reoccuring costs unlike computers.
The reason why it's sold for that price is because there is a huge demand, and people will easily buy it and it will go out of stock faster than you can say that it's too expensive.
The reality is that you simply aren't the target audience anymore.
This is just late stage capitalism 101... it doesn't matter if YOU get priced out, there will be a smaller and smaller segment of the population with more and more wealth that won't be priced out. And they will gobble these up like candy bars.
There is nothing late stage capitalistic about this, it's simply supply and demand.
And it's really not a small population, there is massive growth in the mid to high mid income earning population who are into the hobby of computers. Computers are one of the cheapest hobbies around, for that segment.
These are basically luxury goods; it's not a necessary purchase.
Someone was harassing me about my computer hardware once - I pointed out the computer on my desk cost less than their four wheeler and I use it a lot more frequently.
Shrug! :)
There isn't going to be a middle class let alone an upper middle class. This entire presentation has been about AI. They care about corporations not "upper middle class" consumers.
"The reality is that you simply aren't the target audience anymore." - I don't play games so in that regard I was never their target. I do develop AI models as a hobby. I can afford it but to what end? Where do you see humans in the future?
??? What are you trying to say. Nvidia is a profit-making company, they care about making a profit, yes. Why are you saying it like it's a new revelation?
Do you understand how the free market works? For something to be priced in a way, it needs to sell. If it sells at said price point, that means it's targeted towards a consumer base that already exists.
I don't quite understand the, where do you see humans in the future, how does that have any relevance with the market segments of computer hardware.
I see a future where humans become more efficient, technologies that remove bottlenecks from human progress, and make the capable more capable. I'm a libertarian myself, so my principle is the free market, and consolidating work to efficient technology is going to make things more efficient, and in turn make more money.
More money will be flowing towards skilled labor, and less for menial labor. It's like the industrial revolution, but with technology. I'm all for it.
I didn’t say anything about Nvidia not being for profit. They get waaaaaaay more money from corporate right now then they are from the general consumer. Clearly they are going to sell to anyone or everyone that will buy. My point is it’s not about games anymore as you could clearly see in the presentation. So who is going to invest in their $10k gaming rig?
Literally the entire presentation after the first like 5 minutes for gaming was about AI replacing knowledge and skills jobs with more efficient replacements using AI. In the future the only people working will be the people who feed the ai. If 90% of people are replaced who will buy the gpu? There is no UBI so where is the “money” going to come from?
You seem to have gone completely out of scope and topic of the discussion about the rtx blackwell series and are asking about, philosophical and political aspect of the future for some reason?
Nothing you said has any relevance to any time scope for the rtx blackwell series.
Your questions about population control and money are very political/philosophical questions, and I don't think are relevant to this discussion. In general, life on earth follows the survival of the fittest principle, so I assume Darwinism will just keep happening, you are thinking too deep into it.
1) This conversation is about the price of the 5090. Affordability IS within scope.
2) You steered the conversation this way by talking about the "upper middle class" (which is like 17% of the population), implied I couldn't afford it, tried to make a case that it's more affordable to buy a card then enjoying other hobbies (which is true) and talking down to me like "You need to understand"...No I do... I own a 4090 for that reason. Things are only getting more expensive.
3) You can still buy 30XX series cards that came out 4 years ago. Companies are talking about flipping the switch on replacing the workforce at all levels of work in the next 2 years which is within a 4 year time table in which Blackwell will still be on the market.
4)If you're talking about video games you don't need a 5090 (the topic at hand). If you did what would be the point of playing if only the top income earners can play with you? How do these top income earners have so much free time to play games?
I loved how the crows went dead quiet when he said that like he was expecting some cheering about spending 10k on their desktop.
How absolutely out of touch.
25% increase in price for 33% increase in memory, with 100% increase in bandwidth. 3090 to 4090 was same memory for a $100 increase and very small bandwidth increase. I don't see the problem.
Yep, I don't think people are digesting how big a jump the 1.8TB/sec is. This is going to be a card that ages well over time due to the huge bandwidth. In multi-gpu setups it will really shine. I would have really loved 48GB VRAM just like everyone else, but the lack of it doesn't make this card a bad deal.
Well the 70 series equaling the last flagship is traditionally what happens, was true for the 970 and the 1070 and the 4070 super. In this case I think there’s some AI trick to it, probably dlss4 vs dlss3 but there’s no way it’s dlss turned off.
They are using a transformer architecture for the new DLSS, not sure how much of a difference it will make but maybe takes into account older frame history and further away features
the point being that old-school write opengl once and run it on a gpu is dead. People were up in arms over AMD running an llm w/ more than 24gb vram. This is the same chicanery.
it's because 5090 straddles professional/Quadro territory. it's like asking why red cameras cost so much more than regular DSLR. nvda has no competitors and decided that free compute improvements every generation are no more.
still, I suspect 25% higher cost for 80% higher bandwidth and 33% more memory is perfectly fine for most buyers. honestly even if there were a 5090 with 96gb vram it probably wouldn't run a 72b model at more than 15 tps. I can't imagine much value in that vs a cheaper 32gb variant and I suspect Nvidia understands this too.
models 30b and below are reasonable for most consumers to serve themselves, 70 and above and you're buying a lot of hardware while having low tps due to compute.
of course, I think the sweet spot is 48gb VRAM - 2x3090 or m1/APU is the sweet spot imo. pessimistic prediction is that local llama makes increasingly less sense considering the economies of scale Blackwell and beyond are gonna bring. selfhosting an moe for example is completely not worth the cost unless you have very high utilization in your use case. ofc, considering test time scaling you could achieve high utilization, but low tps means super long wait times for test time scaling. Blackwell racks will have 1000x 4090 bandwidth lol, 5x lower tps might not be a big deal if it's 1 min vs 12s, but it's a big deal if it's 5 min vs 1 or 25 vs 5
4090 was underpriced for pretty much it's entire run as a response to the overpricing mistake Nvidia made with 3090Ti. We'll see if 5090 lives up to it.
I'm really curious how many 4090 gets sold as used though. I guess I'm not the only one considering keeping the 4090 for running secondary AI models (eg. VLM/LLM next to an image/video diffusion model), and buying the 5090 to take it's primary place in the system.
Germany's main tech portal heise.de mentions the following MSRP:
5090: €2329
5080: €1169
5070Ti: €879
5070: €649
I assume these would include the German VAT of 19%.
The US doesn't have a VAT. They have sales taxes which differ from state to state (and community), that's why US prices are always listed without the additional sales tax. Although while for example in Germany VAT is 19% and in the Netherlands 21% the max US sales tax is around 9-10%. So when you buy that 5090 in San Francisco you'll end up paying $2172 and not $1999.
**Should I wait for RTX 50 series laptops to come out before buying?**
Currently, I want to buy a new laptop for my learning journey in AI/ML. Specifically, I need it for running models locally, on edge devices, and for Small Language Models. I'm planning to buy the Lenovo Legion Y7000 2024 with a Core i7-13650HX, 24GB of RAM, 512GB of storage, and an RTX 4060 with 8GB of VRAM. The 15.6'' display has a 144Hz refresh rate. This laptop is within my budget. Should I wait and save up for a laptop with an RTX 50 series GPU? My current laptop is seven years old, and I can barely run models larger than 1B parameters.
https://preview.redd.it/5v3bh5hmaibe1.png?width=1976&format=png&auto=webp&s=1f8b0de403a1fef772e5a1e122f64536760648e1
The 5070 vs 4090 comparison is a complete scam in every sense.
the way Jensen talked about it on stage made it sound like the 5070 was totally gonna blow 4090 out of the water. Clearly given the stats from their own website, the 5000 series is barely an upgrade. All that's really changed is the memory and DLSS. Performance wise it sounds like it might help with gaming , but beyond that, the 4090 comparison to the 5070 is still ridiculous.
So, essentially if one already owns a 3060 and all they care about is AI then they really have no reason to get any of those. I mean I'm sorry but why isn't 5060 16GB? Why do they still insist SO much on making vram unaffordable?
I love how NVIDIA manages to hide their real performance year by year. I would love to know what are the actual TFLOPS for each spec of the 5090's.
Also, from a hardware perspective: the fact that they managed to pull out a 2-slot GPU card (apparently with double the power) and the 4090's were 3.5-slots makes me wonder if they fucked up last design or they realized that sales for AI could get better if they fit it in 2-slot.
(note: the 2-slot makes them better to fit in some servers. Now wait for some Chinese OEMs will design some blower version and the perfect cost-power GPU it's ready)
At this point we should be happy then didn't hike the price even more considering Jensen can charge pretty much anything he wants for the high end cards since he has 0 adversaries there.
Wait as long as possible? Pay to use megacorporations' GPUs instead? The RAM on these are not good enough anyway, and the performance per price will be abysmal. Hopefully AMD ups their game and people start to support their video cards for AI better. I always sort by performance per price when looking for a video card (3070 currently, so I'll wait a long time).
Time to upgrade my computer. Get rid of the old right?
https://preview.redd.it/qa29zae87ibe1.jpeg?width=680&format=pjpg&auto=webp&s=7e0376dbcd23be2a54b82a50871e9e0fddae125c
24GB 5090 = Trash card, not buying
Was honestly planning to buy 2x 5090 32GB for $5000 or even $6000.
Now I'll only buy a used 4090 for $1200 from whatever chump is retarded enough to upgrade to this shit patty.
[https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/](https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/)
32GB for RTX 5090, 16GB for 5080 and 5070ti, and 12GB for the 5070.
I also thought, if my 1070ti dies today, that is not a bad price for double the vram, probably a 1000x faster, and modern gaming features and it can fit bigger models compared to my 1070t, and, the price is comparable to what i paid back then.
We'll have to see if 5070 has 4090 level performance for gaming as Jensen claims and how good DLSS 4 is. For our usecases in local ML world, the best value will always be used flagships like 3090 and P40.
You've been reading too much internet economic takes, that's not how markups work, especially for computer parts. Markups are caused by supply shortages or added marketable utility. Nvidia's favorite partner EVGA decided to quit instead of live with Nvidia imposed low margins or try to compete with premium prices they know was doomed.
If you really want to doom about markups the SKU to look at is 5090 which still is stationed as a prosumer card with heavy demand from China gray market shippers to be turned into datacenter inference cards. That will possibly suffer demand spikes overwhelming supply leaving room for markups.
I will be messaging you in 1 month on [**2025-02-07 04:05:59 UTC**](http://www.wolframalpha.com/input/?i=2025-02-07%2004:05:59%20UTC%20To%20Local%20Time) to remind you of [**this link**](https://www.reddit.com/r/LocalLLaMA/comments/1hvhsm7/rtx_5090_blackwell_official_price/m5tj6t7/?context=3)
[**CLICK THIS LINK**](https://www.reddit.com/message/compose/?to=RemindMeBot&subject=Reminder&message=%5Bhttps%3A%2F%2Fwww.reddit.com%2Fr%2FLocalLLaMA%2Fcomments%2F1hvhsm7%2Frtx_5090_blackwell_official_price%2Fm5tj6t7%2F%5D%0A%0ARemindMe%21%202025-02-07%2004%3A05%3A59%20UTC) to send a PM to also be reminded and to reduce spam.
^(Parent commenter can ) [^(delete this message to hide from others.)](https://www.reddit.com/message/compose/?to=RemindMeBot&subject=Delete%20Comment&message=Delete%21%201hvhsm7)
*****
|[^(Info)](https://www.reddit.com/r/RemindMeBot/comments/e1bko7/remindmebot_info_v21/)|[^(Custom)](https://www.reddit.com/message/compose/?to=RemindMeBot&subject=Reminder&message=%5BLink%20or%20message%20inside%20square%20brackets%5D%0A%0ARemindMe%21%20Time%20period%20here)|[^(Your Reminders)](https://www.reddit.com/message/compose/?to=RemindMeBot&subject=List%20Of%20Reminders&message=MyReminders%21)|[^(Feedback)](https://www.reddit.com/message/compose/?to=Watchful1&subject=RemindMeBot%20Feedback)|
|-|-|-|-|
324 Comments
Biggest_Cans@reddit
Ok_Top9254@reddit
nderstand2grow@reddit
petuman@reddit
Cory123125@reddit
False_Grit@reddit
petuman@reddit
False_Grit@reddit
singinst@reddit
MayorWolf@reddit
OverclockingUnicorn@reddit
MayorWolf@reddit
StyMaar@reddit
Ok_Top9254@reddit
emprahsFury@reddit
yhodda@reddit
RICH_life@reddit
Low-Ad4807@reddit
Kooky-Somewhere-2883@reddit (OP)
Low-Ad4807@reddit
AC1colossus@reddit
davew111@reddit
helgur@reddit
grabber4321@reddit
rc_ym@reddit
Nossa30@reddit
rc_ym@reddit
Nossa30@reddit
rc_ym@reddit
NickCanCode@reddit
guchdog@reddit
fury420@reddit
Nepherpitu@reddit
fury420@reddit
Nepherpitu@reddit
fury420@reddit
Karyo_Ten@reddit
auradragon1@reddit
Ashamed-Selection-33@reddit
animealt46@reddit
Ok_Top9254@reddit
KeinNiemand@reddit
mynamasteph@reddit
krystof24@reddit
Minobull@reddit
krystof24@reddit
mynamasteph@reddit
nderstand2grow@reddit
SexyAlienHotTubWater@reddit
animealt46@reddit
bick_nyers@reddit
SexyAlienHotTubWater@reddit
SexyAlienHotTubWater@reddit
AnimalLibrynation@reddit
SexyAlienHotTubWater@reddit
darth_chewbacca@reddit
emprahsFury@reddit
Sciencebitchs@reddit
one-joule@reddit
darth_chewbacca@reddit
strawboard@reddit
330d@reddit
darth_chewbacca@reddit
330d@reddit
JohnSane@reddit
Secure_Hunter_206@reddit
Derendila@reddit
Secure_Hunter_206@reddit
shirotokov@reddit
Haverespect@reddit
RedTheRobot@reddit
ArsNeph@reddit
Glittering_Mouse_883@reddit
Caffeine_Monster@reddit
ArsNeph@reddit
shirotokov@reddit
ArsNeph@reddit
VermillionBlu@reddit
BangkokPadang@reddit
goldenbluesanta@reddit
shirotokov@reddit
VermillionBlu@reddit
finallyifoundvalidUN@reddit
kuil09@reddit
shirotokov@reddit
Spirited_Example_341@reddit
unlikely_ending@reddit
Salendron2@reddit
animealt46@reddit
decrement--@reddit
No-Refrigerator-1672@reddit
Vb_33@reddit
No-Refrigerator-1672@reddit
AppearanceHeavy6724@reddit
Vb_33@reddit
KeinNiemand@reddit
appakaradi@reddit
decrement--@reddit
nderstand2grow@reddit
Rudradev715@reddit
emteedub@reddit
robertotomas@reddit
No-Refrigerator-1672@reddit
wen_mars@reddit
Karyo_Ten@reddit
wen_mars@reddit
Salendron2@reddit
SexyAlienHotTubWater@reddit
popiazaza@reddit
Longjumping-Trust133@reddit
cinemauser333@reddit
poonDaddy99@reddit
0x01_Tukker@reddit
Kooky-Somewhere-2883@reddit (OP)
0x01_Tukker@reddit
yhodda@reddit
KeinNiemand@reddit
yhodda@reddit
RebornZA@reddit
ronoldwp-5464@reddit
Hunting-Succcubus@reddit
Kooky-Somewhere-2883@reddit (OP)
Hunting-Succcubus@reddit
notanNSAagent89@reddit
Kooky-Somewhere-2883@reddit (OP)
Feisty_Olive_7881@reddit
Kooky-Somewhere-2883@reddit (OP)
sabalatotoololol@reddit
Kooky-Somewhere-2883@reddit (OP)
AnomalyNexus@reddit
Kooky-Somewhere-2883@reddit (OP)
robertotomas@reddit
Kooky-Somewhere-2883@reddit (OP)
saintpart2@reddit
Kooky-Somewhere-2883@reddit (OP)
Electronic-Air5728@reddit
Kooky-Somewhere-2883@reddit (OP)
MayorWolf@reddit
Kooky-Somewhere-2883@reddit (OP)
a_chatbot@reddit
Kooky-Somewhere-2883@reddit (OP)
SteveRD1@reddit
0Ark@reddit
False_Grit@reddit
Eritar@reddit
GoalSquasher@reddit
r0kh0rd@reddit
Baumbauer1@reddit
dandanua@reddit
the_forbidden_won@reddit
DarkJanissary@reddit
swagonflyyyy@reddit
DeltaSqueezer@reddit
PermanentLiminality@reddit
swagonflyyyy@reddit
Erdeem@reddit
DeltaSqueezer@reddit
Kooky-Somewhere-2883@reddit (OP)
Standard-Anybody@reddit
DeltaSqueezer@reddit
Kep0a@reddit
lleti@reddit
nokia7110@reddit
Kooky-Somewhere-2883@reddit (OP)
danielhanchen@reddit
animealt46@reddit
DeltaSqueezer@reddit
Bderken@reddit
danielhanchen@reddit
emprahsFury@reddit
nderstand2grow@reddit
OverclockingUnicorn@reddit
uhuge@reddit
OverclockingUnicorn@reddit
danielhanchen@reddit
Massive_Robot_Cactus@reddit
az226@reddit
RayHell666@reddit
nderstand2grow@reddit
ClearlyCylindrical@reddit
RayHell666@reddit
danielhanchen@reddit
emprahsFury@reddit
ab2377@reddit
LSDemon@reddit
danielhanchen@reddit
saksoz@reddit
__some__guy@reddit
danielhanchen@reddit
__some__guy@reddit
danielhanchen@reddit
BigCompetition1064@reddit
danielhanchen@reddit
BigCompetition1064@reddit
ramzeez88@reddit
muchcharles@reddit
danielhanchen@reddit
L3Niflheim@reddit
xflareon@reddit
loyalekoinu88@reddit
CystralSkye@reddit
PeachScary413@reddit
CystralSkye@reddit
teachersecret@reddit
loyalekoinu88@reddit
CystralSkye@reddit
loyalekoinu88@reddit
CystralSkye@reddit
loyalekoinu88@reddit
mycall@reddit
wordyplayer@reddit
Internal-Comment-533@reddit
darth_chewbacca@reddit
wen_mars@reddit
mxforest@reddit
TimChr78@reddit
jd_3d@reddit
emprahsFury@reddit
ourearsan@reddit
Charuru@reddit
itch-@reddit
iamthewhatt@reddit
Charuru@reddit
xsr21@reddit
muchcharles@reddit
emprahsFury@reddit
Chemical_Mode2736@reddit
programmerChilli@reddit
muchcharles@reddit
visionsmemories@reddit
animealt46@reddit
RazzmatazzReal4129@reddit
carnyzzle@reddit
AnomalyNexus@reddit
inagy@reddit
emprahsFury@reddit
JakoDel@reddit
getmevodka@reddit
Ashamed-Selection-33@reddit
cyberdork@reddit
Kooky-Somewhere-2883@reddit (OP)
RebornZA@reddit
cyberdork@reddit
RebornZA@reddit
Soap_n_Duck@reddit
Herr_Drosselmeyer@reddit
grabber4321@reddit
TimChr78@reddit
popiazaza@reddit
Herr_Drosselmeyer@reddit
grabber4321@reddit
Herr_Drosselmeyer@reddit
NeedsMoreMinerals@reddit
FriedAcid@reddit
TimChr78@reddit
esuil@reddit
FriedAcid@reddit
Kooky-Somewhere-2883@reddit (OP)
getmevodka@reddit
omarc1492@reddit
getmevodka@reddit
roselan@reddit
getmevodka@reddit
sebo3d@reddit
JvalesCrespo3@reddit
segmond@reddit
Kooky-Somewhere-2883@reddit (OP)
roselan@reddit
roselan@reddit
hanzoplsswitch@reddit
KingAndromeda@reddit
Soap_n_Duck@reddit
newdoria88@reddit
thekingdaddy69@reddit
Dax_Thrushbane@reddit
Kooky-Somewhere-2883@reddit (OP)
Turkino@reddit
FuzzzyRam@reddit
Kooky-Somewhere-2883@reddit (OP)
FuzzzyRam@reddit
esuil@reddit
FuzzzyRam@reddit
Kooky-Somewhere-2883@reddit (OP)
LoadingALIAS@reddit
AaronFeng47@reddit
Kooky-Somewhere-2883@reddit (OP)
esuil@reddit
getmevodka@reddit
SithLordRising@reddit
shirotokov@reddit
nderstand2grow@reddit
shirotokov@reddit
Vegetable_Sun_9225@reddit
Kooky-Somewhere-2883@reddit (OP)
huyhoangdrop@reddit
Kooky-Somewhere-2883@reddit (OP)
singinst@reddit
sluuuurp@reddit
emprahsFury@reddit
vincentz42@reddit
mxforest@reddit
VertigoOne1@reddit
VertigoOne1@reddit
animealt46@reddit
LSDemon@reddit
carnyzzle@reddit
darth_chewbacca@reddit
animealt46@reddit
darth_chewbacca@reddit
animealt46@reddit
darth_chewbacca@reddit
animealt46@reddit
RemindMeBot@reddit
Different_Fix_2217@reddit
GrosPoulet33@reddit
AnhedoniaJack@reddit
tommy1691@reddit
animealt46@reddit
Specter_Origin@reddit
mxforest@reddit
Specter_Origin@reddit
domdod@reddit
Jerm2560@reddit
grim-432@reddit