Kitguru : Nvidia to release three GeForce GTX 800 graphics cards this October

[FatherMurphy]

I am going to laugh so hard if this is a 20nm part.

Laughing with joy or laughing with derision? 😕

 

[Mand]

I am going to laugh so hard if this is a 20nm part.

Me too. Everyone is so utterly sure that it's 28nm, when there's absolutely no reason to be certain about it at this point. The only reason was a delay at TSMC that pushed things back by about a month. But, that was picked up by a clickbait site, exaggerated into something that would grab headlines, then quoted by other sites who then quoted each other over and over again until it echo-chambered into something we just "know." We don't know. Even now, we just don't.

Will I eat my words if 880 comes out at 28nm? No, because I'm not claiming it will be 20nm for sure. I'm saying that we have no way of knowing, that historically process shrinks have accompanied major Nvidia generational changes, that the full Maxwell line would have been here much faster if it were on 28nm, and that's entirely plausible that the 750Ti was a one-off test of the new architecture on 28nm before they moved to 20nm.

The decision for which process to build on is not something that is done on the spur of the moment. If they had planned on 20nm for the 800 series, then scrapping that plan and going 28nm less than a year before the expected launch is very unlikely - the contracts with the fabs go out way, way longer than that. If there's a delay, then it's delayed, not abandoned. Not unless the delay is extreme to the point of "we might not be able to fill your order." But that sort of thing is incredibly uncommon.

 

[tviceman]

The 6870 was slower than the 5870. GM104 is obviously not the replacement for the GK110 either. That's why I mentioned it simply as a possibility. I know the 6970 was the 5870 replacement. Also, architectures don't automatically scale up well. Look at Pitcairn to Tahiti.

I'm not saying that's what's going to happen. Just talking points.

It's not valid to compare AMD GPUs when talking specifically about Nvidia GPU's. Case in point, GK106 scaled better in perf/mm^2 than GK107, and GK104 scaled better in perf/mm^2 than GK106. Scaling in Kepler among the graphic-oriented chips was fantastic.

Anyways, I'll say it again: 20% faster than a 780 TI with power consumption very close to GTX 770.

 

[Ajay]

Me too. Everyone is so utterly sure that it's 28nm, when there's absolutely no reason to be certain about it at this point. The only reason was a delay at TSMC that pushed things back by about a month. But, that was picked up by a clickbait site, exaggerated into something that would grab headlines, then quoted by other sites who then quoted each other over and over again until it echo-chambered into something we just "know." We don't know. Even now, we just don't.

Well, it's not impossible. But considering that Qualcomm isn't coming out with 20nm parts till 2015 (http://www.anandtech.com/show/7925/...bit-socs-with-lte-category-67-support-in-2015), it's still highly unlikely that that a lower volume/larger die IDM like Nvidia will be pursuing 20nm this year (and if they were, it would be for AIBs closer to $1k than $500).

TSMC's yields were horrible in the first quarter of this year, assuming the yields became 'acceptable' in the second quarter on means that Apple sucked up every last wafer (unless the rumor that Apple's A8 is a 28nm part are true - but I think that is even less likely).

 

[Cellun]

Had to make an account to post my two cents here. I am also waiting for Nvidia to give us the slip and announce 20nm 800 series, but like Ajay, I am very doubtful. News of staying at 28nm hit me right in the heart, as now I believe I will just SLI my 770 if this series isn't anything spectacular and wait until 16nm comes out in 2016 :/
I think TSMC could have made a comeback if it wasn't for Apple snatching most of the work force there.

 

[njdevilsfan87]

Looking at that review, per that AT review the performance gap between the 7970 and 680 either was maintained or widened in favor of the 680 at WHQD resolutions. I also remember the gap between the 580 and 680 staying around 35% at high resolutions such as 2560x1600. Despite the 256 bit and 384 bit bus.

The GTX 680 came with memory clocked 50% higher than the 580. This completely made up for the difference in bus as the 580 and 680 actually ended up having the same exact overall memory bandwidth at stock. So it's not surprising the difference between the two remained the same.

However, I doubt they can repeat it this time. In order to pull of the same with the 880, they would have to clock the memory at 9-10.5ghz to match the memory bandwidth the GK110 cards offer. I think the 880 will perform well on single monitor resolutions, but fall apart in surround (relative to GK110), which isn't its intended market anyway. I don't see the 880 being anywhere near as "impressive" relative to the 780 as the 680 was to the 580. Nvidia did some hyping before Kepler was released but this time around things are much quieter. My guess is 15-20% (with less power consumption) and diminishing gains with resolution.

 

[Grooveriding]

If they stick with 880, perhaps they'll try and push it as a 680 - maybe 780 - replacement and not a 780ti topper. One thing is for certain, it's going to be very underwhelming as a new part compared to the 780ti. They are stuck on 28nm, can't afford current 20nm which appears to solely belong to Apple and companies of the like that can actually afford it today, so it's not going to have the benefit of a new node like the 680 did.

If this was intended to be mid-range gm204 maxwell we can already see they are compensating for the lack of a new node by it having a fairly large die compared to GK104, about 50% larger. Being a realist even about 20% faster than a 780ti would be enough of an achievement on its own taking advantage of Maxwell but still having to use the same 28nm node and using a smaller die than gk110.

We are seeing some of the same fantasies we heard about the 680 before launch with ridiculous price hopes. People were going on rumours the 680 would cost $300 before it launched, now we hear $500 for a part that might be 20% faster than the 780ti, that is selling at $700, and will have no competition. If it manages that performance it will easily cost $700, maybe more.

Of course maybe we are hearing $500 rumours because it is slower or equal to a 780ti, or perhaps being positioned as a 780 replacement, but a more affordable card than the 780ti is.

 

[96Firebird]

One thing is for certain, it's going to be very underwhelming as a new part compared to the 780ti.

Out of curiosity's sake, what price and performance level will the card need to hold in order for you to be content with it?

 

[Grooveriding]

Out of curiosity's sake, what price and performance level will the card need to hold in order for you to be content with it?

For me personally I don't care how much it costs, but it will need to be at least on average 35% faster at high resolutions than a 780ti to make it worth buying. Even that is pretty dismal as a new flagship. Ideally we should see something on the level of 80-100% faster like we saw from GF110 to GK110 when we get a new flagship.

Realistically this card will probably be at best 20% faster.

 

[Cloudfire777]

The only chance of GTX 880 being 20nm is if the leaked die, this one,

is GM200.
It is roughly 430mm2 so it would be smaller than GK110 which is a 530mm2 die. Only way for that to happen is with a different process, aka 20nm.

If it is, then yeah, pretty much everyone will be shocked.

 

[RussianSensation]

Me too. Everyone is so utterly sure that it's 28nm, when there's absolutely no reason to be certain about it at this point. The only reason was a delay at TSMC that pushed things back by about a month. But, that was picked up by a clickbait site, exaggerated into something that would grab headlines, then quoted by other sites who then quoted each other over and over again until it echo-chambered into something we just "know." We don't know. Even now, we just don't.

We don't know 100% but we can use deduction to arrive at a reasonable estimate:

1) The professional analysts that cover the semi-conductor space continue to repeat the same thing over and over for the last 6 months:

"TSMC has been working on the 20nm node since early February but that work was almost exclusively prioritized for Apple and the creation of its 20nm SoCs."

Considering that even Qualcomm won't ship 20nm chips this year despite launching Samsung Alpha F and Note 4 this fall, how do you expect NV to beat Qualcomm and also launch 880 on 20nm in September, essentially beating Apple to launch? Nevermind the fact that Apple places an order for 60-70 million iPhone 6s!

2) A 20nm chip ~ 430mm2 would have far more CUDA cores than the 780Ti. It's going to be severely memory bandwidth and ROP limited at high resolutions to use that many CUDA cores. Also, this contradicts just about every rumor that once 20nm drops, NV will refresh 880 on 20nm in 2015.

3) If NV really had 20nm available to them, they would allocate 20nm wafers for mobile Maxwell GPUs to get the most design wins. Such a product would have unbeatable price/performance. 20nm + new architecture? NV would get 90%+ of mobile dGPU design wins -- a market that's greater in % terms than desktop dGPU and has more growth for NV. There wouldn't be enough 20nm chips lefts over for 860/860Ti/870/880 for the desktop.

4) AMD is the one which tends to release newer nodes before NV. NV tends to be a lot more conservative and squeeze the most out of mature nodes. Since AMD's management already confirmed no 20nm GPUs for 2014, why would NV have one ready for September 2014?

5) To have 20nm 800 line-up ready for September/October, NV would have had to start manufacturing well before that. How would they have been able to do that when TSMC more or less committed 20nm for Apple's SoC this year?

While we can't say with 100% certainty that 20nm is ruled out, there are so many clues that 28nm is what 880 will be manufactured on and 0 rumors or clues that it will be on 20nm. There isn't even 1 supporting article in the last 3 months from Sweclockers, Tom's, Videocardz, Fudzilla, Semiaccurate or any other site that is even hinting at 880 being 20nm.

 

[Mand]

While we can't say with 100% certainty that 20nm is ruled out, there are so many clues that 28nm is what 880 will be manufactured on and 0 rumors or clues that it will be on 20nm. There isn't even 1 supporting article in the last 3 months from Sweclockers, Tom's, Videocardz, Fudzilla, Semiaccurate or any other site that is even hinting at 880 being 20nm.

That's because they all quote each other as sources. None of it has any basis whatsoever in reality.

 

[3DVagabond]

We don't know 100% but we can use deduction to arrive at a reasonable estimate:

1) The professional analysts that cover the semi-conductor space continue to repeat the same thing over and over for the last 6 months:

"TSMC has been working on the 20nm node since early February but that work was almost exclusively prioritized for Apple and the creation of its 20nm SoCs."

Considering that even Qualcomm won't ship 20nm chips this year despite launching Samsung Alpha F and Note 4 this fall, how do you expect NV to beat Qualcomm and also launch 880 on 20nm in September, essentially beating Apple to launch? Nevermind the fact that Apple places an order for 60-70 million iPhone 6s!

2) A 20nm chip ~ 430mm2 would have far more CUDA cores than the 780Ti. It's going to be severely memory bandwidth and ROP limited at high resolutions to use that many CUDA cores. Also, this contradicts just about every rumor that once 20nm drops, NV will refresh 880 on 20nm in 2015.

3) If NV really had 20nm available to them, they would allocate 20nm wafers for mobile Maxwell GPUs to get the most design wins. Such a product would have unbeatable price/performance. 20nm + new architecture? NV would get 90%+ of mobile dGPU design wins -- a market that's greater in % terms than desktop dGPU and has more growth for NV. There wouldn't be enough 20nm chips lefts over for 860/860Ti/870/880 for the desktop.

4) AMD is the one which tends to release newer nodes before NV. NV tends to be a lot more conservative and squeeze the most out of mature nodes. Since AMD's management already confirmed no 20nm GPUs for 2014, why would NV have one ready for September 2014?

5) To have 20nm 800 line-up ready for September/October, NV would have had to start manufacturing well before that. How would they have been able to do that when TSMC more or less committed 20nm for Apple's SoC this year?

While we can't say with 100% certainty that 20nm is ruled out, there are so many clues that 28nm is what 880 will be manufactured on and 0 rumors or clues that it will be on 20nm. There isn't even 1 supporting article in the last 3 months from Sweclockers, Tom's, Videocardz, Fudzilla, Semiaccurate or any other site that is even hinting at 880 being 20nm.

5 reasons why it won't be 20nm.

That's because they all quote each other as sources. None of it has any basis whatsoever in reality.

No reasons why it will be 20nm.

I'd say Russian is basing it on reality. What are you basing this 20nm rumor on?

 

[RussianSensation]

That's because they all quote each other as sources. None of it has any basis whatsoever in reality.

No, they don't.

"Sources of partner manufacturers now report to SweClockers..." ~ Source

Videocardz was one of the first which claimed Maxwell is 28nm first. 750Ti = 28nm. A lot of these sites get info right before next gen cards launch. Usually a lot of them get the preliminary specs incorrect but I don't remember a time when some of these big sites called the wrong node. You know how many clicks they would get if an article popped up that 880 is actually on 20nm? How come none of them is doing a write-up like that this close to launch?

 

[TrulyUncouth]

5 reasons why it won't be 28nm.


No reasons why it will be 20nm.

I'd say Russian is basing it on reality. What are you basing this 20nm rumor on?

I assume you meant 20nm instead of 28nm on that first one.

I could be wrong but I think the point these 20nm hopefuls are trying to make is not that they think it will be on 20nm- just that we don't have conclusive evidence it won't be. So its not really fair to say "What are you basing this rumor on?" They are not necessarily saying it will be 20nm, just that all the "It 100% won't be 20nm" people don't have solid evidence to that effect.

 

[RussianSensation]

That took about 8 months for the 7970GHz to actually reach 680 performance with driver improvements. And the slower and smoother campaign was based on the 680's slower framerate and smoother gameplay. Not based on a myth. It really happened. Then AMD reeled in their drivers to correct the issue. They did a great job.... eventually.

On June 22, 2012, 7970Ghz tied or beat 680 in most reviews:

1) http://techreport.com/review/23150/amd-radeon-hd-7970-ghz-edition/11

2) http://www.computerbase.de/2012-06/test-amd-radeon-hd-7970-ghz-edition/4/

3) http://www.techpowerup.com/reviews/AMD/HD_7970_GHz_Edition/28.html

4) http://www.sweclockers.com/recension/15564-amd-radeon-hd-7970-gigahertz-edition/17#pagehead

5) http://www.hardwarecanucks.com/foru...radeon-hd-7970-3gb-ghz-edition-review-21.html

6) http://www.hardwarecanucks.com/foru...radeon-hd-7970-3gb-ghz-edition-review-21.html

Etc.

It took 8 months for 7970Ghz to claim the last significant win in BF4 where it handily beat the 680 in frame times and FPS. Since Catalyst 12.11 Beta 2, 680 never lead 7970Ghz on average. Besides EVGA Classified and MSI Lightning 680 that could consistently reach 1260-1280mhz Boost, 680 couldn't beat an 1175mhz 7970, and especially not at 1440/1600p.

That's not even the point. 7970 had a 2.5 months uncontested lead over 580, then when 680 launched, in 3 months, 7970Ghz reclaimed the performance crown. By the time 7970Ghz edition started to drop in retail in July, it came with free game bundles and retail prices were $469-489, undercutting 680 2GB and easily undercutting 680 4GB. Now the 680 2GB is compromised in games like Watch Dogs but 7970 OC keeps on trucking.

---

The situation with 880 is different since NV is launching first. That's why it's so hard to predict the price vs. performance standing. 780Ti with rebates is dropping to $600. 680 beat 580 by 33-35% and undercut the 580 3GB version by $50. 880 could easily launch at $550 and beat 780Ti by 25-30% and it wouldn't be that shocking considering 680 did even better than that against the 580. If 780 is selling for $450-500 and it's a 384-bit 561mm2 die, a 28nm 430mm2 880 made on 256-bit die at $550 would rake in huge profit margins for NV.

 

[3DVagabond]

I assume you meant 20nm instead of 28nm on that first one.

I could be wrong but I think the point these 20nm hopefuls are trying to make is not that they think it will be on 20nm- just that we don't have conclusive evidence it won't be. So its not really fair to say "What are you basing this rumor on?" They are not necessarily saying it will be 20nm, just that all the "It 100% won't be 20nm" people don't have solid evidence to that effect.

Thanks for letting me know about the typo.

That's not what I take from it. Saying that the 28nm predictions, "None of it has any basis whatsoever in reality." is a bit stronger than that. As Russian showed, there are lots of real reasons 28nm is being predicted.

 

[piesquared]

Meh, couldn't care less about Maxwell. nv is spinning dollar/perf which basically says that performance is lacking, much like the Maxwell that's already on the market.
Bring on AMD's next cards please. 🙂 Now where is that Tonga thread? Oh yeah those 2 pages of posts are still being reviewed... 🙄

Infraction issued for moderator callout.
-- stahlhart

 

[CrazyElf]

The logical thing to do is to wait until AMD launches it's 390 and 390X (or whatever their are called). The rumor is they are bringing HBM early on. We could see a situation like the 4850/4870 launch where Nvidia is forced to lower prices (higher than expected performance due to efficient die strategy and early GDDR5 adoption). Right now, Nvidia is milking. $1,000 Titan Black, $3,000 Titan Z (does anyone even buy that thing?). Even the 780Ti is not a good price😛erformance.

It's looking like it will all be 28nm for at least 1 more generation. I've heard there won't be a 20nm high power process at TSMC and that they will go straight to 16nm FinFET for high power. AMD also has the option to transfer at least some production to Global Foundries.

 

[AtenRa]

GM204 vs Hawaii (R9 290X). I bet on this one.

GM204 will be faster at 1080p and 1440p, Hawaii and GK110 will be faster at 1600p and up.

GM204 (GTX880 ??) will be released at $499, R9 290X will fall to $379 and Tonga to be released at $249. GTX-780 will be EOL and GTX-780Ti will be at $649.

Nostradamus is back 😛 :biggrin:

 

[CakeMonster]

I have the Asus DCIIU 780 GTX now, which I'm extremely happy with because of the fact that its virtually silent. I have a feeling that because of the need to push more performance out of 28nm, any new chip will need much noisier cooling. The 290* run on high temps, and any revision of those or a new NV chip I have a hard time imagining running as cool as my older 28nm chip. I hope to be proven wrong though....

 

[3DVagabond]

I have the Asus DCIIU 780 GTX now, which I'm extremely happy with because of the fact that its virtually silent. I have a feeling that because of the need to push more performance out of 28nm, any new chip will need much noisier cooling. The 290* run on high temps, and any revision of those or a new NV chip I have a hard time imagining running as cool as my older 28nm chip. I hope to be proven wrong though....

GM204 is supposed to be more efficient than GK110, so it should be easier to cool.

 

[rgallant]

The logical thing to do is to wait until AMD launches it's 390 and 390X (or whatever their are called). The rumor is they are bringing HBM early on. We could see a situation like the 4850/4870 launch where Nvidia is forced to lower prices (higher than expected performance due to efficient die strategy and early GDDR5 adoption). Right now, Nvidia is milking. $1,000 Titan Black, $3,000 Titan Z (does anyone even buy that thing?). Even the 780Ti is not a good price😛erformance.

It's looking like it will all be 28nm for at least 1 more generation. I've heard there won't be a 20nm high power process at TSMC and that they will go straight to 16nm FinFET for high power. AMD also has the option to transfer at least some production to Global Foundries.

anyone know the state of 20nm from gf for gpu's and could they be making amd's 20 nm gpu's at this point in time ?

 

[3DVagabond]

anyone know the state of 20nm from gf for gpu's and could they be making amd's 20 nm gpu's at this point in time ?

Liza Su (sp?) said that GloFo was making GPU's now for AMD, but she also said we would only see 28nm this year.

 

[Silverforce11]

There's no need to be concern that GM204 will obsolete Gk110 SKUs based on price, that is what its meant to do, same 28nm node, if NV can squeeze more dies per wafer out of it, they can effectively replace the entire GTX700 series. That's what they are aiming to do with this 3X GM204 SKU release.

Now the question is will we ever see big maxwell on 28nm or is that put on hold for 20nm/16nm?