NVIDIA Kepler GPU Speculation thread

[shompa]

I dunno. So far there is no prototype chip for the kepler. That would indicate the chances of a new card in 6 months is near zero. It also just recently taped out.

From chatting with someone in the know, it can take a signifigant amount of time from tapeout > prototype, and then the prototype has to be validated. And then its generally 6 months after the validated prototype that cards are released.

How fast Kepler will be out depends a large part if Nvidia orders risk wafers or not (like they did with Fermi). Risk wafers can nock of 3 month of development time. But even with Risk wafers: A perfect, no respin tapeout that have never happened in history, it would take 3 month before Nvidia can deliver. With 1 respin we are in 6 month range.

BTW. We are seeing an Apple effect here. Apple will refresh its computer line in April. Nvidia is desperate to deliver silicon to Apple. Therefore Nvidia have prioritized the mid range GPUs. According to another website Nvidia had test silicon of these chips in September and will release the GPUs in late Q1.

 

[Rifter]

And I didn't say it was him. Obviously you can't read yourself.

actually you did.

You're posting the unconfirmed rumours that were floating around for a few months already. Creating a new redundant thread with old rumours the day new tech launches. Err I smell biasiness. ()

Saying that below his quote means you are talking about him, who else would the "you're posting the unconfimed rumours" at the beggining mean other than him, the only one quoted.

 

[AtenRa]

Alright. I'll save this one. It's meaningless to be here anymore. See you in april.

Thread title = NVIDIA Kepler GPU Speculation thread

My estimation is,

 

[RussianSensation]

Alright. I'll save this one. It's meaningless to be here anymore. See you in april.

It won't be very wise of me to start laughing at this specs now, because if this turns out to be true I'll make a fool of myself. But but but I'll make sure to bring this up if you are wrong.

:whiste:

speculation

[spek-yuh-ley-shuh

n] http://dictionary.reference.com/help/luna/Spell_pron_key.htmlconjectural consideration of a matter; conjecture or surmise: a report based on speculation rather than facts.

If you have no interest in this topic, please stop harassing every poster.

 

[nsavop]

The only Kepler that has a remote chance of launching within the next few months is gk104. But with NV confirming its due for late Q2 however, its chance of launching soon is next to nil.

gk104, 768 cuda cores, a node shrink gtx580. On specs it has potential for 50% gains, but we all know that it never scales perfectly with increased resources. Assume good gains of 80%, it should be ~30-35% faster than gtx580. This is the only part NV has to compete with 7970, because gk100/110 is even further away.

I had said in the other post, if NV needed a respin to fix gk104 delaying it to late Q2, and its just a "simple" die shrink of gtx580.. do you really think they won't need respins for gk100/110, a chip with 1024 cores and new kepler features bloating the die size well beyond gtx580's 520mm2, all on a new node at TSMC. Doesn't take much logic to figure kepler high-end isn't going to be here anytime soon.

If they drop the hotclock design, its even more harder to predict performance. ATM, gtx580's 512 cuda cores run at what, 1.7ghz? If gk104's 768 cores run at 1ghz, it would be slower. It needs to run at ~1.2ghz to reach parity, and >1.3ghz to claim slightly faster than gtx580.. not impossible, but initial rumors had gk104 as "slightly faster than gtx580".

Maybe your right, all these different part numbers some maybe kepler some maybe a fermi die shrink with release dates from Q1 to the end of 2012. Kepler could be having issues causing Nvidia to panic and do a quick shrink of fermi to get them by. Who knows

 

[shompa]

I wonder if TSMC has a love/hate relationship with Nvidia. The bean counters see the volume and do a happy dance, the engineers see the transistor count and head for the bottle.

I'm more interested to see what Nvidia is going to be doing on the compute side of things myself.

TSMC does not care. They charge per wafer, not per good die.

But I do wonder how fast 28nm will ramp up and if they will have problem like they did with 40nm. In Q4 less then 3% if TSMC revenue was from 28nm.

We probably have Nvidia/Apple and AMD fighting for 28nm wafers. If TSMC scored the A6 contract they need to deliver an insane amount of wafers to Apple. (Apple needs about 20 million A6 per month when iPad/Iphone uses it and if Apple released the rumored ARM laptop at least 25 million/month)

 

[Saico]

actually you did.



Saying that below his quote means you are talking about him, who else would the "you're posting the unconfimed rumours" at the beggining mean other than him, the only one quoted.

No I didn't. Who else? Ofc the topic starter. If you're too dumb to understand that by "The guy who started this thread" I mean TOPIC STARTER then it's not my fault.

 

[shompa]

Hence the reason for a speculation thread. I think we can all agree on one thing, Nvidia needs to get Kepler out ASAP, or at least have refreshed Fermi parts coming out shorty.

When is Nvidia's next earnings call? That is probably when we will get some concrete-ish info.

I think you might be imaging things.

Why do Nvidia need to get out Kepler ASAP?
AMDs latest card is 15-20% faster then GTX 580 and costs a lot more. DX 11.1 is irrelevant since there won't be any DX11.1 games for at least 1+ year.

The most interesting thing about AMDs graphic cards is that they finally have destroyed Dual Link DVI limit of 8 gig/bandwidth for a monitor. Finally we technically can have higher resolution then 1920x1080 120Hz / 2560x1600 60 Hz. The 2560x1600 monitor was released in Aug 2004!

This high speed HDMI is something that Nvidia needs to deliver when higher resolution displays starts to manufacture Q2. Nvidia will release mid marked Kepler for this market in time.

 

[blastingcap]

TSMC does not care. They charge per wafer, not per good die.

But I do wonder how fast 28nm will ramp up and if they will have problem like they did with 40nm. In Q4 less then 3% if TSMC revenue was from 28nm.

We probably have Nvidia/Apple and AMD fighting for 28nm wafers. If TSMC scored the A6 contract they need to deliver an insane amount of wafers to Apple. (Apple needs about 20 million A6 per month when iPad/Iphone uses it and if Apple released the rumored ARM laptop at least 25 million/month)

Bad news if true. Maybe AMD wanted to price the 7970 but couldn't without destroying their margins, thanks to Apple buying up all the 28nm capacity? If so, Nvidia would probably follow suit and we get only a modest bump up in price/performance. Damn Apple.

 

[RussianSensation]

Maybe your right, all these different part numbers some maybe kepler some maybe a fermi die shrink with release dates from Q1 to the end of 2012. Kepler could be having issues causing Nvidia to panic and do a quick shrink of fermi to get them by. Who knows

That's possible. 4Gamer.net has a slide that shows no high performing Kepler parts until at least middle of 2012. Again, in stark contrast to TPU's report of GK104 slated for Q1 2012.

 

[OVerLoRDI]

expect AMD to refresh HD7970 shortly without problems (like Q2)

Really? That quick?

If the above road map is correct it looks like we have some time to wait for some shiny high end NVidia parts. For myself I won't be upgrading, till NVidia shows their cards or a 7970 refresh comes out. I will be recommending the 7970 as an upgrade for those running 5870s who want more performance now (and have the cash).

 

[boxleitnerb]

Why does anyone believe 4gamers? They were wrong before on so many occasions with roadmaps just like this.

 

[BD231]

Im not going to count on a doubling in cores being that nvidia cores are so much faster than amds, and I bet they go conservative, its not like theyre hurting for a competitive part.

Theyll milk the fermi, soon to be rebadged kelper for all its worth.

 

[Arkadrel]

512-bit bus

How fast is NVidia gonna get that to go?

Assumeing 1375mhz on the ram, around 350 GB/s.

Im not going to count on a doubling in cores being that nvidia cores are so much faster than amds, and I bet they go conservative, its not like theyre hurting for a competitive part.

^ this.

its not possible for them to double the memory bandwidth, with just 512bit bus, thus they ll lose out on peformance due to memory if they made a chip that big.

Then there is the matter of TSMC ~1.6x (optimistic) gain.
This is from a TSMC spokes person, giveing a "optimistic" outlook on gains.

520mm^2 on 40nm -----> 520/1.6 = 325mm^2 on 28nm.

Since a 520, has 512cores, a 1024 core, would likely be twice as big.

Okay.... 325mm^2 doubled up = 650mm^2.


Then you add in new features, that might take up more space ect..... yeah...
1024 core fermi for 680 is definately out of the picture.



That said Nvidia doesnt need that big a increase to beat the 7970.
Im guessing 768 cores or so..... which should put it around 20-25% faster than the 7970.

 

[tviceman]

BTW. We are seeing an Apple effect here. Apple will refresh its computer line in April. Nvidia is desperate to deliver silicon to Apple. Therefore Nvidia have prioritized the mid range GPUs.

This sounds logical, but Apple also uses fairly high end GPU's in some of their iMac's (they used an hd7970m), so Nvidia will need to deliver more than just the lowest end Kepler's to Apple. GK104 is going to be Kepler's GF110 - Nvidia's best selling 500 series card AND the highest end mobile GPU chip used by Nvidia on 40nm as well. I expect GK104 to operate in the same capacity.

 

[dust]

Really? That quick?

If the above road map is correct it looks like we have some time to wait for some shiny high end NVidia parts. For myself I won't be upgrading, till NVidia shows their cards or a 7970 refresh comes out. I will be recommending the 7970 as an upgrade for those running 5870s who want more performance now (and have the cash).

Yeah RS's post reminded me of the comment Apoppin made in his preview, that 7970 may be short lived:hmm:

 

[Idontcare]

(This is a Moderator Post, not a member post)

Well, you're not any better than anyone from the hordes of "7970 60% better than 580" fanboys.

Stripping off "rambus and 1024 cores" from GTX780 we'll get another 10-20% faster than 7970 for 200$ more, 6 months later. Get real.

Do I need to repeat my post?

I'm just really surprised by your attitude. Yesterday you were bashing and moking 7970 for not being as good as the speculations suggested it to be. And now you're posting your own speculations and want people to believe in them. :whiste:

You're posting the unconfirmed rumours that were floating around for a few months already. Creating a new redundant thread with old rumours the day new tech launches. Err I smell biasiness. ()



And they didn't confirm it's a Kepler. GK100 to be precise. Thye have some card in the works to counter 7970 but everyone is saying the GK100 is not anywhere close to prototype.

All I know, is that some Nvidia fanboys are really upset with AMD having a nice lead with their new GPU's and try really hard to spoil this launch by lowering real-world perfomance numbers for it and posting false rumours about Nvidia's future offerings. I'm not pointing my finger at anyone. No. Just FYI.

And personally for you RS - I suggest learning to use forum search to find all threads marked by KEPLER to read all old rumours you've been reposting here. The closest one is not even a page away.

Ah, yeah learn to read other people posts too because you keep refering to TPU and 3D as "reputable sources" while they were wrong all the way about 7970.

Hey, don't pull a RussianSensation on us. The link is 5 posts above your own.

And I didn't say it was him. Obviously you can't read yourself.

Alright. I'll save this one. It's meaningless to be here anymore. See you in april.

It won't be very wise of me to start laughing at this specs now, because if this turns out to be true I'll make a fool of myself. But but but I'll make sure to bring this up if you are wrong.

No I didn't. Who else? Ofc the topic starter. If you're too dumb to understand that by "The guy who started this thread" I mean TOPIC STARTER then it's not my fault.

Saico,

You are relatively new here in these forums so you may not be entirely up to speed on the expectations of how you are to carry yourself and the forum etiquette you need to employ in these technical forums.

Allow me to introduce you to Graham's Hierarchy of Disagreement (abbreviated here, hit the link for full text replete with examples):

DH0. Name-calling.

This is the lowest form of disagreement, and probably also the most common.

DH1. Ad Hominem.

An ad hominem attack is not quite as weak as mere name-calling. Saying that an author lacks the authority to write about a topic is a variant of ad hominem—and a particularly useless sort, because good ideas often come from outsiders. The question is whether the author is correct or not. If his lack of authority caused him to make mistakes, point those out. And if it didn't, it's not a problem.

DH2. Responding to Tone.

The next level up we start to see responses to the writing, rather than the writer. The lowest form of these is to disagree with the author's tone.
DH3. Contradiction.

In this stage we finally get responses to what was said, rather than how or by whom. The lowest form of response to an argument is simply to state the opposing case, with little or no supporting evidence.

DH4. Counterargument.

At level 4 we reach the first form of convincing disagreement: counterargument. Forms up to this point can usually be ignored as proving nothing. Counterargument might prove something. The problem is, it's hard to say exactly what.

DH5. Refutation.

The most convincing form of disagreement is refutation. It's also the rarest, because it's the most work. Indeed, the disagreement hierarchy forms a kind of pyramid, in the sense that the higher you go the fewer instances you find.

DH6. Refuting the Central Point.

The force of a refutation depends on what you refute. The most powerful form of disagreement is to refute someone's central point.

I have gone to the effort of capturing this Hierarchy of Disagreement in a way that shows you what is the minimum level of acceptable posting in the various subforums here on the AnandTech Forums:

Notice that posts which contain elements of DH0, DH1, and DH2 are unacceptable in the technical forums (which includes Video Cards and Graphics, aka VC&G).

And why do we have guidelines to prevent the employment of methods of disagreement that span name-calling, ad-hominem attacks, and responding to tone?

Because it is needless (there are better, higher DH#, ways to express your disagreement), it is inflammatory, and perhaps most crucially it is ineffective to your own end-goal of actually convincing members of the community that your position is correct.

Now we have a way of classifying forms of disagreement. What good is it? One thing the disagreement hierarchy doesn't give us is a way of picking a winner. DH levels merely describe the form of a statement, not whether it's correct. A DH6 response could still be completely mistaken.

But while DH levels don't set a lower bound on the convincingness of a reply, they do set an upper bound. A DH6 response might be unconvincing, but a DH2 or lower response is always unconvincing.

The most obvious advantage of classifying the forms of disagreement is that it will help people to evaluate what they read. In particular, it will help them to see through intellectually dishonest arguments.

We all need you, Saico, to clean up your act in this forum. You are needlessly combative, your posts are laced with unproductive hyperbole and rhetoric, and the majority of your posts (in fact every single one of them in this thread to date) are in violation of the posting guidelines.

Lastly, if you or anyone else in the community wishes to discuss this specific Moderator Post I urge you to do so by way of creating a thread in the Moderator Discussion subforum rather than posting here so that we avoid derailing this thread (as well as risking committing a mod-challenge) from its intended purpose of speculating on Nvidia's Kepler GPU.

Administrator Idontcare

 

[blackened23]

Why do Nvidia need to get out Kepler ASAP?
AMDs latest card is 15-20% faster then GTX 580 and costs a lot more.

Say what? Let me take some time to correct your errors.

GTX 580 3gb MSRP (EVGA stock version): 589.99 per http://www.evga.com/products/moreInfo.asp?pn=03G-P3-1584-AR&family=GeForce 500 Series Family

7970 3gb MSRP: 549.99

 

[nsavop]

Say what? Let me take some time to correct your errors.

GTX 580 3gb MSRP (EVGA stock version): 589.99 per http://www.evga.com/products/moreInf...eries%20Family

7970 3gb MSRP: 549.99

Nvidia should be dropping the price of the 580 to $350-400 come Jan.9 or shortly there after, if not they risk the 580 becoming irrelevant.

What interests me is does the 560ti and 570 follow suit, those cards might also get a welcomed price cut.

 

[blackened23]

Nvidia should be dropping the price of the 580 to $350-400 come Jan.9 or shortly there after, if not they risk the 580 becoming irrelevant.

I hope so, then everyone wins :thumbsup:

 

[MrTeal]

I'm sure everyone has seen this already, but I'll post it up here just to add it to the thread.

Source: http://legitreviews.com/news/12046/

Not sure how much faith I'd put in this, but given the doubling of performance I would think this is referring to the 1048 core part.

 

[Olikan]

I'm sure everyone has seen this already, but I'll post it up here just to add it to the thread.
Not sure how much faith I'd put in this, but given the doubling of performance I would think this is referring to the 1048 core part.

to double the performance, they will have to double the bandwidth,
it's impossible even using 512bit-rate

 

[Genx87]

GK100 = GTX590 performance
GK104 = ~7970 performance

 

[Flipped Gazelle]

TSMC does not care. They charge per wafer, not per good die.

But I do wonder how fast 28nm will ramp up and if they will have problem like they did with 40nm. In Q4 less then 3% if TSMC revenue was from 28nm.

We probably have Nvidia/Apple and AMD fighting for 28nm wafers. If TSMC scored the A6 contract they need to deliver an insane amount of wafers to Apple. (Apple needs about 20 million A6 per month when iPad/Iphone uses it and if Apple released the rumored ARM laptop at least 25 million/month)

In their contract, I'd imagine there is language pertaining to the % of good dies. Also, TSMC is not the only game in town, and poor performance on their part would encourage clients to go else, or perhaps even invest in their own fabs.

 

[MrTeal]

to double the performance, they will have to double the bandwidth,
it's impossible even using 512bit-rate

What, the memory bandwidth? That would only be true if GF110 performance was being gated exclusively by memory. Even then, 384-bit@4GHz gives 192GB/sec. If NVIDIA could get the memory frequency up to 5.5GHz like AMD is able to, that would by 352GB/sec. It's not a doubling, but it's pretty close.