NVIDIA 4000 Series

[Armageus]

Colorful confirms GeForce RTX 4070 Ti SUPER with AD102 GPUs - VideoCardz.com

AD102 confirmed in RTX 4070 Ti SUPER Colorful has introduced and confirmed the new RTX 4070 Ti with AD102 GPU. No more wondering or pixel peeping. Colorful has officially confirmed their latest RTX 4070 Ti SUPER cards, dubbed “V2,” feature the NVIDIA AD102 graphics processor. This is the first...

videocardz.com

Colorful too.

@BornEvil

RTX 4070 Ti Super GPUs Based on AD102

HW News - Microsoft Keeps Ruining Windows, Valve Lawsuit, '4090' in the 4070 Ti Super

Sponsor: Thermal Grizzly Aeronaut on Amazon https://geni.us/e8Oq & Hydronaut (Amazon) https://geni.us/hOQrBAbMicrosoft is shoving automatic OneDrive uploads ...

youtu.be

Bonkers how cut down the AD102 has had to be to hit that level - over half the die disabled from the full theoretical 18432 core part.

	Model	Die	CUDA Cores	Memory Bus (bits)	Bandwidth (GB/s)	VRAM (GB)	TDP (W)	Price
Desktop	4060	AD107	3072	128	272	8	115	$299
Desktop	4060Ti	AD106	4352	128	288	8/16	160	$399/$499
Theoretical	AD106 Max	AD106	4608	128
Desktop	4070	AD104	5888	192	504	12	200	$599
Desktop	4070 Super	AD104	7168	192	504	12	220	$599
Desktop	4070 Ti	AD104	7680	192	504	12	285	$799
Theoretical	AD104 Max	AD104	7680	192
Desktop	4070 Ti Super	AD103	8448	256	672	16	285	$799
Desktop	4080	AD103	9728	256	761.8	16	320	$1199
Desktop	4080 Super	AD103	10240	256	736	16	320	$999
Theoretical	AD103 Max	AD103	10240	256
Desktop	4090D (China Only)	AD102	14592	384	1008	24	425
Desktop	4090	AD102	16384	384	1008	24	450	$1599
Theoretical	AD102 Max	AD102	18432​	384

One big question remaining for the 40xx series... Where is the 4080Ti ?

Also with how cut down the 4090 is I suspect that should have been the real 4080Ti and a real 4090 with more cores or a fully unlocked chip, but they sold that last gen with a 2% increase in cores for $2000 as the 3090Ti FE and that didn't go down well as we remember.

Why is the 4080Ti still MIA...? We have had an 80Ti since the 780Ti.. Before that all full chip 80 class cards were just xx80's and no Ti, then also the Titans that played the so called top cards that were either cut down chips or full chips as it depended on Nvidia's mood.. As we remember with the Titan X and then Xp nonsense.

4080 Super was the fully enabled AD103, so a Ti would have to have come from the AD102. I guess it just depends whether bits could be disabled in a "balanced" enough way to delivery better than 4080 Performance.
The stumbling block possibly being the 384 bit memory bus - leaving that enabled = 24GB and so is too attractive against the flagship 4090, but if it has to be disabled back to 256 bit, perhaps the other sections end up to weak to be worthwhile?

 

[Blackjack Davy]

Given the £400 difference would you go for a 4090 over a 4080?

Because you don't like having your pockets emptied even further by a man in a leather jacket? Just a hunch.

 

[JediFragger]

Then yes, shall be avoiding amd

**** off, if they managed to create all your wishes and cure cancer at the same time you still wouldn't touch them with a 10ft pole

We know you too well by now!!

 

[TNA]

**** off, if they managed to create all your wishes and cure cancer at the same time you still wouldn't touch them with a 10ft pole

We know you too well by now!!

Is he that far gone? Lol

You do know his previous gpu was an AMD one? And the one before that and the one before that.

You can see all his pro Radeon posts for years on the forum

 

[JediFragger]

You do know his previous gpu was an AMD one? And the one before that and the one before that.

I know, just ribbing him a touch

He's right ofc, AMD just have too much to do in order to catch-up it ain't gonna happen.

I got this XTX for about £650 which was priced about right imo and can live without certain features so wth, tbf it's been solid and fast for the money

 

[Nexus18]

**** off, if they managed to create all your wishes and cure cancer at the same time you still wouldn't touch them with a 10ft pole

We know you too well by now!!

Maybe if they give me a reach around I might consider them then

But in all seriousness (and I know you are joking although some on here will sadly believe that is the case anyway ), I had amd dGPU's right from the 3850 to vega 56 (have only ever owned 3 nvidia gpus, 3rd being my current 3080) and owned amd ryzen 2600, 5600x and then 5800x3d (with no intentions on changing this anytime soon) so it's simple, don't make **** products and I'll buy

But tbh, amd have put me of them with the way they have gone about things over the last 4 years or so i.e. this we are premium brand nonsense when they don't match the market leading brand, playing the sympathy card that time with how devs shouldn't be supporting closed source methods and instead to go with their open source solution (even though they're last to the market with an inferior solution....) and other things like making out they are the white knights for gaming scene and so on.

 

[Nexus18]

He's right ofc, AMD just have too much to do in order to catch-up it ain't gonna happen.

Funny thing is people think nvidia were way ahead years back with more defining features with things like physx but they weren't really, physx is the only real thing they had and imo, it was pretty crap, not to mention, very few games with such features, main ones being the batman arkham series and metro. AMD back then were actually better because of the advantage they had with dx 12 when it was coming about as well as mantle (although in very few games but made a real nice difference when paired with my weaker cpu the 2600 especially in BF 4, which was my main game). Gsync was probably the main and most important advantage as nvidia had the market to themselves with an issue free experience for a good couple of years until freesync became more common and had improved VRR ranges along with the introduction of low frame compensation and not having black screen/flickering issues. It's only really in the last 2-3 years nvidia have really pushed ahead now.

 

[Tinek]

Best move was the ones who got 3080 for £650 and have saved money by skipping the scalped 3080s and the ridiculous 3090 jenson no lube edition and avoided ada entirely

I never got my preorder for 3080, even though I ordered just when they got released. months waiting and I just gave up and went for 6800xt from ocuk instead (for close to MSRP). I got 3080Ti eventually too but that was way later and didn't last long till I swapped is for 4090.

4090 was the only gpu to make sense at launch and even if folks hold onto it and skip 50xx entirely, they'll still have an incredible gpu but as shown, the ones who buy such gpus like that are often the ones who are more likely to upgrade every new gen to the new flagship

Nvidia would have to show something truly amazing for me to bother with 5k series. And I'm quite confident they won't. It feels like 4090 was too good for their own future plans. As Jensen said, it was too cheap. ;p

Couldn't care less on the number/tier of gpu now, bang per buck is what is important.

Always true, unless someone just do not care. I actually looked at that carefully before pulling the trigger.

 

[GoogalyMoogaly]

Maybe if they give me a reach around I might consider them then

So you'd be OK with being bent over by AMD too, just as long as they sweeten the deal

But tbh, amd have put me of them with the way they have gone about things over the last 4 years or so i.e. this we are premium brand nonsense when they don't match the market leading brand, playing the sympathy card that time with how devs shouldn't be supporting closed source methods and instead to go with their open source solution (even though they're last to the market with an inferior solution....) and other things like making out they are the white knights for gaming scene and so on.

Now I have no proof that it went/goes down this way, but the feeling I always get with the AMD vs Nvidia tech stuff is that Nvidia produce something, like Hairworks, which is blackbox and then they are active in getting devs to use it and help devs to implement it. Then AMD produce TressFX and go "Here's the github README". And then when a dev shows interest in using it AMD goes "READ THE ******* README!!! IT'S ON GITHUB FFS!" and if the devs ask for any assistance it's just "README. GITHUB."
And then we get lots of games with Hairworks in and 1 with TressFX and then AMD and certain corners of the internet complain about Nvidia's tactics.

Again, no proof, it's just how I imagine it playing out.

 

[Nexus18]

Now I have no proof that it went/goes down this way, but the feeling I always get with the AMD vs Nvidia tech stuff is that Nvidia produce something, like Hairworks, which is blackbox and then they are active in getting devs to use it and help devs to implement it. Then AMD produce TressFX and go "Here's the github README". And then when a dev shows interest in using it AMD goes "READ THE ******* README!!! IT'S ON GITHUB FFS!" and if the devs ask for any assistance it's just "README. GITHUB."
And then we get lots of games with Hairworks in and 1 with TressFX and then AMD and certain corners of the internet complain about Nvidia's tactics.

Again, no proof, it's just how I imagine it playing out.

Based on my own experience working with open and closed source solutions and the customer experience, you're probably not far from the truth when it comes to nvidia and amds solutions, which is why at work, we are often more than happy to go closed source and even pay extra to get a good experience with support as opposed to "go and figure it out".

That and the fact that amds open source tech such as the fireworks, tressfx never really takes of outside of a handful of games kind of backs that up.

Also, from reading on their github page for FSR where devs asked questions on a few things such as the updating of FSR to newer versions, they even responded with something along the lines of "it's entirely up to the dev as to how to incorporate FSR". Roy and the guy who took over from him even basically said how they like to work and why they favour open source as it allows them to be more hands of and let the community work with it i.e. they can do the over the fence approach, which there is nothing wrong with (especially if you have a smaller/lesser skilled team, heck if it weren't for mantle, we probably wouldn't have vulkan and dx 12 would have taken longer so this is where open source does pay of i.e. when it involves into something bigger that the industry will use) but as shown by lack of and slow uptake of said tech. it's a problem if you're wanting to get your technology adopted quickly and at scale, which obviously when you have such a low market share in the first place is probably not the best strategy....

In terms of technology, I think freesync and fsr are probably amds biggest successes and that's largely thanks to them having consoles where such tech could and will be very beneficial but we have seen the issue with slow uptake here too regarding FSR on xbx and ps 5 (sony are even looking to do their own solution which kind of says it all imo)....

Essentially AMD need to be careful they don't fall too far behind and that sony/microsoft won't ever be enticed by nvidias offerings.....

 

[Rob_B]

Why are they cutting off half a 102 to make these? Surely yields can't be that bad? And are these companies releasing this to suggest they're better than a non 102 version?

 

[Murphy]

Why are they cutting off half a 102 to make these?

They're not, they're fusing off the parts that don't work.

Surely yields can't be that bad?

If there's enough failed 4090 dies to bother spinning up another SKU then you'd have to say yes they're that bad.

And are these companies releasing this to suggest they're better than a non 102 version?

No, the physical size of the die is the same as that used in 4090's, in fact they would've gone into 4090's if all the parts of the die worked as they should've.

The defect could be that they weren't stable at 4090 clock speeds, it could be a defect SM, a defective Tensor, RT, CUDA core, or anything. They all go into a pile and if it's cost effective to spin up another SKU they'll blow some fuses on the die so functionally it's the same as a 4070.

 

[Rob_B]

I mean I know they don't literally cut them in half

More surprised there are that many bad 102s (I guess they've been building up) and I'm suspecting that manufacturers are releasing this info so some buyers think "4090 die is better than a 'normal' 4070 Ti Super"

 

[Tetras]

Why are they cutting off half a 102 to make these? Surely yields can't be that bad? And are these companies releasing this to suggest they're better than a non 102 version?

It seems weird when they can sell GPUs for AI faster than they can make them, maybe there are future sanctions that will limit their sales, or they're just shifting stock in preparation for blackwell.

Has anyone seen how they perform? I wonder if the huge die would help with cooling, but have a bit higher power consumption since disabling stuff doesn't ever seem to be 100% effective.

 

[Rroff]

Surely yields can't be that bad?

TSMC has had several partial shutdowns over the 4000 series production cycle so possible some batch(es) aren't up to the normal quality level - with at least one of those events online wafers were variously damaged or had higher than normal defects - though I hadn't heard of it affecting the lines nVidia stuff was on.

 

[MrMarvelous]

Tin foil hat here - I think they're fusing them off as there's a load of unsold 4090 chips. To not devalue the 4090, these have been turned into a 4070Ti to get rid as it's really the only card that sells en masse to consumers.

 

[mrk]

So you'd be OK with being bent over by AMD too, just as long as they sweeten the deal


Now I have no proof that it went/goes down this way, but the feeling I always get with the AMD vs Nvidia tech stuff is that Nvidia produce something, like Hairworks, which is blackbox and then they are active in getting devs to use it and help devs to implement it. Then AMD produce TressFX and go "Here's the github README". And then when a dev shows interest in using it AMD goes "READ THE ******* README!!! IT'S ON GITHUB FFS!" and if the devs ask for any assistance it's just "README. GITHUB."
And then we get lots of games with Hairworks in and 1 with TressFX and then AMD and certain corners of the internet complain about Nvidia's tactics.

Again, no proof, it's just how I imagine it playing out.

 

[Murphy]

I mean I know they don't literally cut them in half

Ahh, sorry. I can take thing a little to literally sometimes.

Surely yields can't be that bad?

Without knowing exact numbers, amount of 4090's made/sold, how many of these new 4070's there will be, it's hard to know for sure but you have to assume it's bad enough to make spinning up a new SKU cost effective.

It could also be that yields aren't that bad it's just the profit margins are so good that it's worth spinning up this new SKU.

And are these companies releasing this to suggest they're better than a non 102 version?

The marketing departments sure will.

 

[brostradamus]

amd is pivoting to software development it seems, reinventing the wheel... while nvidia takes even bigger strides, soon their manufacturing advantage over intel is also going to vanish, got to be an interesting time for short-sellers

 

[brostradamus]