NVIDIA reveals die sizes for GB200 Blackwell GPUs: GB202 is 750mm², features 92.2B transistors
Posted by fatso486@reddit | hardware | View on Reddit | 175 comments
DerpSenpai@reddit
so this is to be expected. samish density as last gen, improvements come just from the new architecture
randomkidlol@reddit
the new architectures are good enough that they can get a performance improvement over last gen using less silicon, whereas traditionally you'd only see a performance improvement on roughly the same amount of silicon.
so nvidia does the most profitable thing and upsell lower tier silicon as competitors to a higher tier card from last generation, but mess around the naming a bit. thats how you get rebranded xx60 dies selling for $550 as a 4070.
rabouilethefirst@reddit
It's AMD's fault at this point. If the 9070 XT is trash, it's their fault NVIDIA is doing this. They truly have no competition.
UGMadness@reddit
It's hard for AMD to keep up the pace given how vast the differences in R&D budget there are between the two. Nvidia is swimming in cash.
SuperDuperSkateCrew@reddit
Yup, if AMD ever seriously threatened Nvidia’s position they’ll just throw billions into R&D to overtake AMD.
Last time AMD was relatively competitive with Nvidia’s high end was with their Fury cards and Nvidia responded with the 10 series and AMD hasn’t really come close to touching their halo products. Nvidia has significantly more money now than they did then.
DerpSenpai@reddit
Poor AMD, they just are the chipmaker of this gens consoles, millions of phones that use Exynos and all iGPUs in AMD systems, they had no money to invest but they also can use 10s of billions in Stock Buybacks...
The fact is, Lisa got caught with her pants down and Jensen said "don't mind if i do"
Darkknight1939@reddit
This sounds so bad out of context, lmao.
DerpSenpai@reddit
In Portuguese there's a saying that
"quanto mais prima, mais se lhe arrima"
As in, it doesn't matter if she's a cousin
BigBasket9778@reddit
No, Jensen! Not your cousin! 😱😰
rabouilethefirst@reddit
AMD has had the answers to the test since 2018. AI upscaling was a good idea, it's the only feature missing from their cards that people would actually care about. RT performance is a bonus. They have refused to do anything for years. The RX 7000 series is a disgrace imo, and they shamelessly tried to pretend they could compete with NVIDIA without even giving proper hardware for an AI upscaler.
Now those cards are gonna be worth next to nothing when FSR4 is not backwards compatible.
bob-@reddit
Whenever I see someone say "AI x" I just clock out can't read further than that bullshit marketing term that means nothing
Qesa@reddit
I can definitely most definitely blame AMD for buying ATI then immediately slashing their R&D budget.
ThrowawayusGenerica@reddit
They were still going strong up to GCN, and then decided to refresh the same chips over and over and over again for 5 goddamn generations.
Adromedae@reddit
They weren't that strong even back then.
Since Kepler NVIDIA hasn't had much competition. They haven't had much pressure and have been able to focus on DC. For a few generations now, NVIDIA has been selling their old value tier die SKUs as premium tier, because AMD just doesn't have an answer in the halo/premium tier segments of yore.
Qesa@reddit
Yeah early GCN was the last thing unaffected by the budget cuts and layoffs
DerpSenpai@reddit
unlike AMD, Nvidia has only the dekstop and laptop market, they don't serve the Console (except Nintendo but thats low value) and integrated graphics markets. AMD just took for granted
randomkidlol@reddit
if AMD refuses to get their shit together, then intel is the last hope in building at decent product at a competitive price point.
SuperDuperSkateCrew@reddit
If Intel actually sticks with their dGPU’s I can see them easily stepping up in the next 5 or so years. Their cards are already great performers for the price but are hampered by the infancy of their driver/software support.
I can see them being highly competitive with AMD in that time period.
oomp_@reddit
why do people blame the competition for a businesses greedy pricing. Blame the ones doing the greedy pricing. businesses don't have to be greedy
rabouilethefirst@reddit
Nobody is forcing you to buy an NVIDIA graphics card, you just know that getting the comparable AMD one for even a reduced price will basically lock you out of tons of features and give a worse experience.
I don’t really “blame” either, but NVIDIA has no reason to drop prices, or even really give more performance at this point. Until AMD has an AI upscaler, buying anything but NVIDIA is a foolish thing.
You would think AMD would have taken the hint about AI upscaling being really important earlier, but no, they waited 7 years, and now here we are.
cp5184@reddit
But because customers overwhelmingly choose nvidia, and keep following nvidias price hikes and asking nvidia for more bigger price hikes... Nvidia ends up sitting on a pile of money able to pay for more engineers better engineers, to even get custom exclusive graphics ram.
AMD doesn't have the money or the people to compete at an equal level with nvidia. Nvidia can simply outspend them because consumers refuse to buy AMD.
Consumers will pay more for a 1060 3GB with fewer CUDA cores than a 1060 6GB than they will for a better performing AMD GPU with more vram.
viperabyss@reddit
I mean, AMD has been swimming in cash since they've been eating into Intel's server marketshare. It is AMD who prioritized CPU development over GPU, believing that would be more lucrative.
It's all about opportunity cost.
Starcast@reddit
I feel like this sub vastly overestimates the number of gamers who fuck with graphics settings or even install AMD/NVIDIA software. I'd dare say 90% of us, myself included, couldn't tell the difference between FSR and DLSS in the real world.
Strazdas1@reddit
Most people install the software because thats the default option when installing the driver and most people leave everything on default.
Jaznavav@reddit
You need to be genuinely blind to not notice the difference.
TheElectroPrince@reddit
This is not a "businesses being greedy" problem. This is a capitalism problem.
Capitalism allows greed to flourish to the extremes.
oomp_@reddit
this is a problem is both the system and the business being greedy, even under our broken system they don't have to be greedy but I guess that's harder when you're publicly traded and have parasitic shareholders having voting rights vs the people who do the actual work, the workers. companies need to democratized
TheElectroPrince@reddit
And I'm saying that will never happen under capitalism.
What we need is socialism.
Or at least social democracy.
Muted-Green-2880@reddit
How is the 9070xt trash? If they price it properly it will dominate the midrange. Should beat the 5070ti slightly in raster while costing 40% less if they price it at $499, more likely they'll be stupid and match the 5070 pricing but that's still a good midrange card. The 5080 will only be 15-20% ahead in raster for also half the price lol. Hardly trash
rabouilethefirst@reddit
I don’t think it will be trash, I just think the 7000 series was an epic fail at launch, and I don’t trust AMD yet. They basically ran it back with the 6000 series and priced $100 under NVIDIA.
9070XT is much more promising with an actual AI upscaler.
Muted-Green-2880@reddit
They have a lot of potential with this card and they're under new leadership, I'd have to think they've learned at least something from their last few launches, they actually went backwards in markershare with the 7000 series. They've said themselves that they're targeting markershare.....as to what that means for them remains to be seen but anything over $499 isn't going to be aggressive enough imo. I have a feeling they'll price it at $549. Which is acceptable but won't enough to gain much markershare. They need to undercut the 5070 while beating it easily in raster, they have a winner on their hands if they don't botch the launch pricing lol. They really need to announce asap
Chipay@reddit
Poor NVIDIA, forced to raise their prices because they have no competition /s
NVIDIA is a big boy company, they can decide to fuck over their customers by themselves.
rabouilethefirst@reddit
NVIDIA is caught twiddling their thumbs and inventing new ways to use their uber powerful tensor cores, because they are so bored at the lack of competition. They even kinda sort of dropped prices this gen.
AMD has had the answers to the test since 2018, and STILL has not released a card with AI upscaling. Even sony beat them to the punch with their own custom solution on RDNA type hardware. They are the main problem here.
AMD and their users dig their heads into the sand and screech about how FSR upscaling is better because it works on ALL cards, when the reality is, that is its main WEAKNESS.
Thankfully the 9070XT looks to correct this glaring issue that has been a problem for 7 years, too bad for all the people that thought they should buy AMD for internet points, because those cards are now OBSOLETE.
Meanwhile, NVIDIA is using tensor cores from RTX 2000 series to bring DLSS 4 to everyone. It's been a joke for AMD up until now.
TheElectroPrince@reddit
Radeon is really only surviving with B2B architecture contracts at this point.
bladex1234@reddit
So does every other company when there’s no competition. Good old capitalism baby.
hackenclaw@reddit
It takse ryzen have about 30%-50% more performance at the same price point to crack Intel's mindshare. Even that it took up until Zen 3. Thats 4 generations(including Zen+) price dominance back to back.
If AMD want to break nvidia's market. They need to offer 30-50% more performance for 3 generation back to back.
Darkknight1939@reddit
Because that's the first generation they actually outperformed intel, lol.
Zen and Zen+ were still embarrassing showings for them. The 8700k vs 2700x was brutal, and the 9900k had no rival.
Single core performance is still the most important metric for performance. You obviously need multiform, too.
They brought cheap multithreading performance. They started gaining marketshare with Zen 3 because IPC/Single core performance is the mist important metric.
gomurifle@reddit
They have to dripbfeed the market too. Can't give us too much of a performance bump. 20% to 25% is more than enough.
DerpSenpai@reddit
That was done in the 4000s gen, This gen is the same as 5000 when talking dies. Also, technically, there hasn't been as big of a GPU as the 5090 so it makes sense we are not seeing anything made from it. IT'S A BEAST in size. It only exists because there is a huge demand for professional work with 5090s
Adromedae@reddit
The 5090 exists because the Halo effect is a thing, and it works as a marketing ploy. And also because, there is a market of gamers that can afford it and will purchase them.
DerpSenpai@reddit
The 5090 is far more than gamers. in my old employers building, it's a place where each workstation has a 4090.
Adromedae@reddit
That says more about the scale/solvency of your former employer, than the trends for the x90 tiers.
LinkesAuge@reddit
It says as much about his employer as companies paying for Mac products or an Adobe subscription (or really expensive office chairs).
The price for Nvidia cards is really nothing special in an enterprise context, if anything there are plenty of enterprise hardware products that have far, far worse ratios in regards to price/utility.
Adromedae@reddit
RTX 5090s are not enterprise products. That was the point.
randomkidlol@reddit
the TU102 and GV100 dies were larger than the incoming GB202 die.
there has always been demand for the top tier die in the form of enterprise customers buying the most expensive quadros/titans by the dozens. the xx90 cards are essentially defective quadros with missing SMs and without ECC or any of the software validation or long term software support of a quadro. the issue here is nvidia's raising the prices of quadros because what else are you gonna buy, and now the $2000 nerfed consumer card looks attractive for customers who are not $100mil+ companies.
Adromedae@reddit
FWIW The Geforce line has always been the nerfed Quadros. At least the x90/x80/x70 series.
tukatu0@reddit
There was no arch update. It's just the upgrade was like 6 minor node gens. Going from a snap dragon 865 to sp 8 gen 1 essentially.
I thought there was going to be an m arch update this gen. But no. Not even that either. Literally just more electricity consumption. Ray tracing 20% uplift comes from instead of you fps dropping rrom 100 to 50. It will drop to 60. Viola. 20-30% improvement.
There is the ai stuff. But i do not think any games will use it within the time frame of these cards relevance. *maybe the 5070 acts as if it has 24gb. Well the ps6 might have 32gb mem with 70gb effective. So eh. Good luck ¯\(ツ)/¯
Erikthered00@reddit
Now we just wait for all the YouTube videos where they butcher the wording.
“750 millimeters squared”
No! It’s 750 square millimeters. 750 millimeters squared is (750mm)^2 which is not the same as 750mm^2. It works out to be 562,500mm^2 as they say it and it’s clearly wrong.
Gamers Nexus, LTT, hardware unboxed have all done this regularly.
_Oxygenator_@reddit
nobody fucking cares
Erikthered00@reddit
Oddly antagonistic response.
Anyone with a science or engineering background generally does. If they want to be seen as knowledgeable, then amateur errors like this really undercut it.
_Oxygenator_@reddit
Oh yes mister expert. Tell us all about the correct way to communicate. Never mind the fact that absolutely no one was misunderstanding what was being said.
Erikthered00@reddit
Again, hostile.
Seriously, I think you need to look at how you talk and think about people.
bexamous@reddit
562,500mm, 562,500mm2 would be "750 square millimeters squared".
dirthurts@reddit
They are REALLY holding back on die size on everything below the 90 series. Clearly trying to push everyone to spend the maximum amount. It makes the value of the mid tier chips look bad (because they are).
I miss the days with the 70 and 80 series got most of the die and good amounts of VRAM.
Noreng@reddit
I don't think it's a matter of limiting the die size so much as Blackwell not scaling all that well beyond the GB203 die size for gaming loads. The GB203 is 50% larger than the GB205, and performs 50% better. The GB202 is 100% larger than the GB203, but only performs 50% better.
dirthurts@reddit
The end user doesn't care about scaling. just about performance, and they're bleeding us for it.
Far_Success_1896@reddit
Did you think that you were going to get 4090 performance for $999 or something?
dirthurts@reddit
They did say that, to be fair. Actually they said 4090 for 4070 ti prices. They lied. You're fine with it?
Far_Success_1896@reddit
I'm getting a 5080 yea. Are you?
dirthurts@reddit
Heck no.
Far_Success_1896@reddit
Great. I'm sure your next card will be marketing free and ethically sourced.
Strazdas1@reddit
Must include free range generated frames. None of that artificial pixel inclusion.
Noreng@reddit
Say Nvidia decided to make a GB202.5 then, it'd be about 600 mm^2 in die size. What about performance?
Logically it'd be in-between the 5080 and 5090, so somewhere in the 20-25% faster than 5080, with 24GB VRAM on a 384-bit bus. About 1500 USD in cost. That's a "5080 Ti" for you.
Now, who would buy this? Once you're at this price point, why not go full send for the 5090? Or just remain at the 5080 for 999 USD?
rabouilethefirst@reddit
Not offering a $1k card that has 4090 performance is a fail imo. I don't care how they do it, they aren't offering much this gen.
Noreng@reddit
Nvidia is offering a slight upgrade for the same price, that's better than Ada was at launch at least. Even if it's not really changing much.
rabouilethefirst@reddit
But people want more performance. ADA was what I wanted. A new node, massive performance gains, redesigned architecture, more features. They could have just done price slashes on ADA and the people would have been happy, that's how far behind AMD is.
Strazdas1@reddit
It does not matter what people want though. Only matters what people are willing to pay for.
DerpSenpai@reddit
4nm is just revised 5nm, there's no new node. Also 3nm is INSANELY costly for GPUs and simply not available yet. You want fairy tales and what we have is reality.
SoTOP@reddit
Making GB203 bigger would have made much more sense then having another die in between, something ~450mm2 with 4090 beating performance using 320bit 20GB vram. Would differentiate 5080 and 4080 much more. Of course it's questionable if that would be more profitable, especially with no AMD high end card around.
Noreng@reddit
And then the 5080 is at least $1199 MSRP, while the 5070 Ti is $999 MSRP.
I don't think people would be more happy that way.
BuchMaister@reddit
I think the gap between 999$ and 1999$ is big enough to insert at least another card, probably there will be one at the super refresh cycle. A lot of people who can spend 1200-1500$ will prefer to go down to 5080 than spending 500-800$ for 5090. You can also look at that from the other side, as someone who planed to spend 1000$ might be tempted to take step up in price bracket to 1300$ or 1400$ if performance improvement is good enough.
ResponsibleJudge3172@reddit
A card only 15% faster than a 5080. There is no point
sharkyzarous@reddit
More like 999 to 2999, the demand for 5090 will be so HUGE that thing will be scalped to beyond our imaginations, and i think partner cards will be over 1999 even with zero scalping. As you said a 1300-1500 price point will be quite tempting while in bite the bullet mode.
Noreng@reddit
It'll be a 448-bit card for 1599 with a 10% bump over the 4090
BuchMaister@reddit
or 1200$-1300$ for 4090 like specs\performance is also possible. There are a lot of possible configurations, and they can even insert 2 cards in that price gap if they wanted. I just wonder why they did not add more die between the GB203 and GB202 a die around the 550 mm\^2 would have made sense, but who knows.
Noreng@reddit
There's no die coming in-between the GB202 and GB203, it wouldn't make sense to design one. Rasterization scaling from more SMs is almost dead beyond the GB203, which is why the GB202 is twice the size.
BuchMaister@reddit
I know there is no die coming between GB202 to GB203. l still don't know what the exact difference in performance in perf between 5080 and 5090 but if it's somewhere around 40-50% than a die half way through would give 20-30% improvement. Maybe they can save lot of defective dies enough to make product (or two) in the middle, which is why they decided not to make another die. In any case there is definitely space for cards between 5080 and 5090 both in terms of specs and price.
Noreng@reddit
Jensen talked about the 4090 as an investment, this is why the 5080 is slower, and the 5090 is more expensive.
BuchMaister@reddit
Nice to know that my 4090 is an investment lol. The real reason is wafer prices going up and them wanting to keep high profit margin.
Noreng@reddit
I'm pretty sure Nvidia is paying less per GB203 than AD103, though it's probably not enough to offset the cost of GDDR7 and the improved cooler.
dirthurts@reddit
I think it's a little strange to assume another 500 USD is just this trivial amount that most people just just toss towards a bigger card.
I could have spent 2k more on my car and upgraded it to the higher model, I didn't. It's a similar philosophy.
4060 vs 4060 ti is proof that saving a few bucks is a common tradeoff that is readily taken.
Noreng@reddit
It's a matter of making the increase in performance noticeable. There's no logical reason for why Nvidia would make a chip in-between the GB203 and GB202, therefore it doesn't exist. It would likely increase prices of all cards due to added design costs.
4060 vs 4060 Ti is a much different case. Because the 4060 lowers the entry price, and cutting costs matters a lot more on that end of the spectrum. Nvidia still makes more money by selling to data centers, but by cheaping out on die size and PCB with the 4060, they managed to create a card that was significantly cheaper than the 4060 Ti while performing surprisingly well.
Vb_33@reddit
These cards are cheaper than the 40 series were. You're not gonna get cheaper and much more performant in the same gen anymore. It's either modest gains and modest price (50 series) or high gains and high price (40 series). Brace your butt for the 60 series they'll be on N3 and will bring larger gains at a cost..
DYMAXIONman@reddit
The 50 series will be the worst generational uplift by Nvidia in a long time. It's even worse from what we got from the 1000 series to the 2000 series.
LinkesAuge@reddit
Funny to bring up the 2000 series when it aged extremely well, especially thanks to the same (software/AI) features that are once again being criticized.
It feels like people still haven't understood that you can't just look at "raw raster performance" anymore and judge the quality of cards just on that, it's short sighted and ignores where technology is actually going.
DYMAXIONman@reddit
It was considered a bad uplift compared other generations but the 4000 ->5000 and the 3000 -> 4000 were way worse.
dirthurts@reddit
They're cheaper than the launch models, which were then quickly replaced with the super variants which were a price to performance correction/reaction. We should be looking at those prices/spec, which makes these look like an even worse value.
You're letting them play you like a fiddle. They're pointing their finger exactly where they want you to look and you're falling for it.
Noobpwnerr@reddit
Nah he's just being realistic no one is getting played lol. Nvidia is basically competing with themselves at this point, so their pricing is whatever they think will be low enough to nudge consumers to upgrade their 3070/3080 to the current gen.
ResponsibleJudge3172@reddit
We also know that TSMC is charging 30% more for 3nm.
999*1.3 is about $1300 for MSRP at the same margins 5080 vs a 3nm 5080
Vb_33@reddit
Ok then I'm sure AMD is going to bring the mad raster gains we used to have at modest prices. Let me guess the 9070 is going to be a 3070 equivalent and will therefore match the 4090 like the 3070 matched the 2080ti. And it'll launch at a $500 MSRP. Let's see how well that goes.
dirthurts@reddit
You are very strange.
cstar1996@reddit
The supers released over halfway through the lifecycle of the generation. That is not “quickly replaced” by any metric dude.
rabouilethefirst@reddit
"Hey look, we made the FPS counter say 200!"
dirthurts@reddit
Pure manipulation.
"why does my game look nice but play like crap?
*whispers at the boot screen "Nvidia!"
CryptikTwo@reddit
No they’re not, the 5070 and the 5080 are, 5070ti and 5090 are more expensive. And the whole price is fixed to performance is bullshit as well, jensen can go fuck himself with that one. More than happy to take my money to AMD or intel at this point.
Vb_33@reddit
The 5070ti is $50 cheaper than the 4070ti was.
CryptikTwo@reddit
You’re right I thought it was $849 for some reason.
996forever@reddit
Of course the end user doesn't care, they're just explaining diminishing returns might be why nV isn't willing to use the biggest die on sub-flagship cards.
dirthurts@reddit
That's just the marketing reason. The real reason is to up-sale the user. It's all about wringing the customer and don't let them convince you otherwise.
ThrowAwayP3nonxl@reddit
I see you're new to this whole capitalism thing
dirthurts@reddit
Not new. Just not complacent.
996forever@reddit
Marketing? Nobody irl even knows what a die size is. Die size is not part of marketing.
randomkidlol@reddit
die size is part of the BOM for each card. lower the BOM, up the naming tier, crank up marketing bullshit with "AI features" and "less is more", maximize increase profit margins.
996forever@reddit
Yes the "AI features" and "less is more" are the bits designed to market and not the die size itself.
dirthurts@reddit
Literally in the article.
996forever@reddit
Yes, they released this piece of specification details. Like they would its memory bus size, transistor count, and compute throughput. Do they think a gaming audience any less or more likely to buy this product because of such details? No.
DiCePWNeD@reddit
It's likely both reasons.
Is there an engineering reason for their technical decision in consumer Blackwell? Sure.
But you'd have to be delusional if you think Nvidia hadn't thought about the business strategy of having this design either, and it's likely the primary focus for them
FinalBase7@reddit
Diminishing returns start hitting hard at the high end but not at the low end, the 4080 is only 33% larger than the 4070 but performs 60% better, Nvidia has been using node shrinks to significantly reduce die sizes of mid range and stagnate core counts instead of keeping the die the same size and filling it with more cores for better performance, the 4060, 4060Ti and 4070 are all a disgrace, they reduced the amount of cores on some and kept the others the same as last gen and used the efficiency improvements to just push clocks as high as possible to make up for the stagnation.
4070 super has 20% more cores and performs 15% better than 4070, with basically no clock speed or memory speed differences, the scaling is still extremely good at mid range and low end
Qesa@reddit
The 4070 is significantly cut down whereas the 4080 isn't. Comparing full dies (4070 ti vs 4080 super) AD103 is ~30% faster
FinalBase7@reddit
I mean that's still really good scaling?
Dey_EatDaPooPoo@reddit
Sure, but you were making a completely misleading comparison when you didn't need to. But yes, the scaling is very good up until GB203.
Noreng@reddit
Yes, there's room for a GB204
Alarchy@reddit
Gaming GPUs are the scraps now. The big money (10-50x profit) is on enterprise. Their compute segment (enterprise) is now 3x the revenue of graphics (gaming). Nvidia is taking a massive profit hit on using wafers for gaming (vs. for GB200), but it's to keep a halo product for PR dominance.
Equivalent-Bet-8771@reddit
Gaming GPUs also get their products out to developers so they can have an enterprise market AT ALL. Businesses don't make things happen it's the employed humans that do it.
Alarchy@reddit
I was responding to the OP on why the die sizes are so small for everything except the 5090. You don't need a 4090/5090 to putz around with dev.
Equivalent-Bet-8771@reddit
Yeah you do bned a 5090. Those RTX cores are weak as hell for what developers are doing with them. If anything the 5090 isn't enough. Have you seen how shit raytracing looks? And that's with complex software tricks to improve it visually.
SERIVUBSEV@reddit
Gaming GPUs are on same die as before 4N, while GB200 is on TSMC 4NP.
You think Nvidia is the type of company to leave money on the table to help out gamers?
Morningst4r@reddit
NVIDIA will likely be giving more attention to gaming than its share of profit might suggest because it’s a safe market that’s consistent. That’s why they did things like LHR cards while 95% of the GPUs they were selling were used for crypto.
Vb_33@reddit
Why do people keep spitting this nonsense in every thread?
Alarchy@reddit
You're right, it's more complicated than my flippant soundbite comment.
Nvidia's gaming segment built the company, they're not going to abandon it. But in 2023 data center exploded, and is their fastest growing (and largest by far) revenue stream - as a publicly traded company it's their fiduciary duty to make money hand over fist, and that's the segment they can do it in, not gaming. In just two years* they've grown their revenue 6x, on the extreme demand for data center. The world is desperate for Nvidia compute products, and Nvidia can't produce them fast enough. They're heavily supply constrained on both HBM and packaging and predict they will be until at least 2026; if they weren't then they'd be fulfilling the year+ backlog** of Blackwell on priority. They are taking a massive prospective profit hit using wafers for anything outside the compute segment.
Releasing Blackwell consumer GPUs that don't require HBM makes sense, releasing several huge die size GPUs (the OP's comment pointing out they don't) doesn't make sense. Same as with Ada, it will be one big die, low-volume chip (including a future Quadro variant) for solidly keeping the performance crown in gaming and workstation (halo effect is real), at their usual ~30% improvement over previous gen, and the bulk of the gaming chips they'll sell (and high volume) will be extremely expensive for their comparatively small size. They're not going to outright abandon consumers, and there's still money to be made (just not nearly as much), but their priority is undeniably producing as many data-center (Hopper and Blackwell) chips as they possibly can.
People lament that Nvidia consumer GPUs get so crippled if they're not the xx90 now, but this is why. It will be the same going forward, unless something drastic changes (AI/datacenter crashes, AMD or Intel have a massive tech leap over Nvidia, etc.). Data center is the cash cow, and it makes sense to go all-in on it as much as they possibly can vs. the slowing consumer space.
TL;DR - gaming getting the "scraps".
panckage@reddit
To be fair, didn't Nvidia cut down orders from TSMC? I thought there was more than enough supply of chips for them to use.
Alarchy@reddit
Not exactly. They cut down CoWoS-S orders (Hopper, early Blackwell), because they're shifting to CoWoS-L (rest of Blackwell) - still with TSMC.
tukatu0@reddit
Too many f""" people in this website.
mulletarian@reddit
Because it keeps being repeated without being refuted
III-V@reddit
They're not supply constrained on the wafer side. They don't use HBM on their consumer cards, so there's no packaging shortage either.
burnish-flatland@reddit
Don't worry they don't, hbm is the bottleneck
mckirkus@reddit
My kid has a 3060 12GB and the 4080 is 16GB. They reeeaaaally don't want people using these for AI.
SmokingPuffin@reddit
5080 has a pretty normally sized die for the tier. Smaller than 3080 and 2080, but those were on lame processes. Bigger than 1080. Same size as 4080.
The story is that 5090 is absolutely huge, not that 5080 is weirdly small. In turn, that's about big demand for AI workloads driving big volumes of big parts.
MrMPFR@reddit
Mate it depends on the generations x80 tier die have been anything from \~300ish to 628mm\^2.
Small dies with Kepler, Maxwell, Pascal and Lovelace, big dies with Tesla, Fermi, Big Kepler, Turing and Ampere.
5090 is the second largest consumer die after 2080 TI. A pointless product for gaming, core scaling for rasterized workloads is just terrible, which 30, 40 and 50 series have shown.
Sorry to say it folks but a 5080 TI will not make any sense. Just not worth it for NVIDIA.
SmokingPuffin@reddit
I don't think 5080 Ti makes sense either. Historically, the x80 Ti has been nearly same core counts as the x90/Titan, but less memory. This GB202 design doesn't really yield an interesting product with that strucuture.
Nvidia might do it if they need a price cut by another name, I guess.
dirthurts@reddit
O.K. Hol up.
You admit you're aware they're cutting down the die sizes and charging us more...and you're not seeing the problem? The 80 series are literally lower class cards, re-branded, and being upsold with frame gen. You all learned nothing from the last gen? They even re-named the 4080 and cut the price to the 4070 ti when people complained. Now you're fine with it? They own you.
SmokingPuffin@reddit
I don't agree that Nvidia is cutting down die sizes and charging more. 5080 is same size as 4080. 5070 Ti is now a cut GB203 (83% of 377mm2) rather than a full AD104 for 4070 Ti (294mm2). 5070 is now a full GB205 (263mm2) while 4070 was a cut AD104 (77% of 294mm2). Then the prices are all down $50-$200 gen on gen.
I don't know what it is I'm supposed to be "fine with". Nvidia can release whatever products they like. I will buy them if I like. I didn't complain when the 4080 and 4070 Ti were released. I see that as futile. I just didn't buy either.
panckage@reddit
Press F for the amazingly priced 3060ti MSRP that you couldn't actually purchase lol
Far_Success_1896@reddit
They mightve been able to do that but you would have closer to 4090/5090 pricing on the 80 cards.
They decided to go less performance and go cheaper which probably works better for most gamers. There aren't that many who can take advantage of a 5090 anyway which barely has many use cases for gaming workloads.
fatso486@reddit (OP)
Isnt it odd that the transistor density reduced a bit compared to the ADA ( at
122.9MTr/mm2, while the AD102 is 124.9MTr/mm2.)
Zednot123@reddit
SRAM has high density, memory controllers have low density. Change the ratio and you change density.
bazhvn@reddit
Isn't SRAM low density (aka most troublesome scaling down for most of node jumps)?
Qesa@reddit
No. It has scaled less than logic, but that is logic catching up rather than SRAM falling behind.
Adromedae@reddit
That is not correct.
SRAM libraries have traditionally always lagged behind dynamic logic libs/cells in terms of density and scaling per node advance.
UsernameAvaylable@reddit
But not in absolte density.
Adromedae@reddit
What do you mean by that?
Strazdas1@reddit
SRAM is denser, but scaling is slower so logic is catching up. However in absolute density SRAM is still denser than logic.
III-V@reddit
It's the opposite. SRAM has effectively stopped scaling.
Cute-Pomegranate-966@reddit
Negative it scaled again on tsmc 2nm
Strazdas1@reddit
isnt that more because of GAA and backside power delivery more than actually density inprovement in the library?
DerpSenpai@reddit
and it willl scale again on CFETs and you will be able to stack SRAM the same way you stack NAND Cells.
Cute-Pomegranate-966@reddit
Glorious. Now if only we can make the actual transistor structures smaller we'd have massive shrink lol
Equivalent-Bet-8771@reddit
Getting there. These structures are approaching dozens of atoms in size already.
wintrmt3@reddit
No they aren't, 2nm is means it's equivalent density to 2nm planar transistors according to the marketing department.
Equivalent-Bet-8771@reddit
Excellent news we need that stacked SRAM.
DerpSenpai@reddit
yeah i doubt we will see too many stacked SRAM cells for L1 or L2 but L3 caches are about to get monstrous in the next 10 years.
Equivalent-Bet-8771@reddit
My body is ready.
Adromedae@reddit
It "scaled" but not at the same rate as dynamic logic.
SRAM and DRAM are effectively the canaries in the coal mine in terms of starting to hit walls of transistor and circuit element scalability in CMOS.
Qesa@reddit
It's effectively stopped, but still has a peak density of ~300 MT/mm^(2). Logic is like 180 MT/mm^(2) for N4 and 220 on N3. Not that anywhere near peak is ever realised, but it's easier to get closer to it for SRAM without brownouts.
But also, logic is only still scaling due to fin depopulation, and with N3 they have a 1-1 cell so that's also hit a wall. And the move to GAA will also benefit SRAM
Strazdas1@reddit
depends on how you count transistors it could mean anything from real improvements to less dummy transistors.
HenryWasBeingHenry@reddit
Wonder if they will do a N3P Blackwell refresh next year. Anyone knows how 4N compares to N3P?
6950@reddit
Nooooo but our lord Jensen Said he is progressing Faster than Moore's Law /s
Zednot123@reddit
not really, since cache amount didn't go up. And you added memory controllers.
SRAM has high density and IMCs have relatively low density. So you you changed the ratio of low/high density areas of the die.
HorrorCranberry1165@reddit
Total costs of 5090:
- 250$ for GPU, assuming 15k$/wafer and 10% yield loss
- 150$ for 32GB GDDR7 - assuming 50% more costly than GDDR6
- 200$ for PCB, cooling, others
- 55$ - R&D and drivers -> 10000 engineers * 150k$/year * 2 years = 3 bln$, 7mln sales desktop GPU per quarter * 8 quarters = 56 mln GPUs, 3 bln$ / 56 mln = 53$ / GPU, not counting GPU sales for laptops
Final price is 650$, MSRP price - 2000$
More you buy - more you save - FOR US
noiserr@reddit
How much is the tape out cost?
Reddia@reddit
I read somewhere gddr7 is 12$/gb atm
Swimming-Low3750@reddit
Nvidia has insane profit margins. This can be seen in their corporate filings!
Bitlovin@reddit
Look around, that’s literally everything. The entire world revolves around increasing revenue growth infinitely, simply turning a profit isn’t enough anymore. The consumer is fucked.
Fr0stCy@reddit
10% yield loss is a miracle and borderline impossible. I’ve seen smaller dies struggle at 50% yield loss.
Thesadisticinventor@reddit
I think 10% yield loss is kinda optimistic for such a large chip tbh, but that shouldn't affect the point you are trying to make. The margins are insane.
HorrorCranberry1165@reddit
Don't think so, GPU specs are lower than real number of existing resources (shadres, different untts, blocks) just to be more tolerant for random partial damages. I think 10% are GPU with more than planned flaws, so they have resources below spec.
JuanElMinero@reddit
The only other time Nvidia consumer cards used a die that large was the Turing with the 2080 ti.
2080 ti was on TSMC 12nm, which was refined TSMC 16nm, right around the time 7nm hit the mass market.
Never before was there a consumer GPU of such size, on the highest performing node available for mass production.
amazingspiderlesbian@reddit
Isn't tsmc 3 in mass production so this is still a gen behind like turing
Reactor-Licker@reddit
Nothing consumer level uses N3 class nodes currently except for Apple. Heck, even Nvidia’s sweet sweet data center GPUs are still on N4.
Zarmazarma@reddit
This release definitely rhymes with the 2000 series. Lots of new, unproven features, with what appears to be a very modest improvement in top end performance (30%, similar to 1080ti -> 2080ti).
randomkidlol@reddit
the 2080ti was the defective version. the full chip was only in the titan RTX, quadro rtx 4000, and quadro rtx 6000.
Adromedae@reddit
True, but those are not consumer SKUs.
MrMPFR@reddit
Indeed this is the second largest consumer die ever. Think Ampere takes third place.
I would bet the BOM for a 5090 is +2-3x that of a 2080 TI. Process node is 2-3x more expensive, almost 3x more memory, +325W higher TDP.
The cost of producing bleeding edge products is getting absurd.
Haematoman@reddit
How the fuck do we even make 92.2 Billion transistors on them. It's fucking insane. Space Magic.
AutoModerator@reddit
Hello fatso486! Please double check that this submission is original reporting and is not an unverified rumor or repost that does not rise to the standards of /r/hardware. If this link is reporting on the work of another site/source or is an unverified rumor, please delete this submission. If this warning is in error, please report this comment and we will remove it.
I am a bot, and this action was performed automatically. Please contact the moderators of this subreddit if you have any questions or concerns.