NVIDIA Pascal Thread

[antihelten]

So for a comparison, aftermarked 980 ti's are getting scores from ~22k to 23k.

http://www.guru3d.com/articles_page...0_ti_extreme_gaming_waterforce_review,23.html.

How do we know this are the "1070" and not the "1080" btw ?

We don't know that this is the 1070, we don't even know if this is GP104, it might very well just be a fake number using a 980 Ti.

Btw the numbers are from the same guy who made these videos

So it seems that this unknown card is ~4% faster then the fastest aftermarked 980 TI.

But we still don't know if this is the "1070", the "1080" or even the "1080ti".

huh? 20625 is only about 2% faster than a stock 980 Ti (20172), it's nowhere near the fastest aftermarket 980 Ti (~23000).

edit: nvm, I guess you were referring to the numbers 96Firebird posted, my bad.

 

[tential]

So it seems that this unknown card is ~4% faster then the fastest aftermarked 980 TI.

But we still don't know if this is the "1070", the "1080" or even the "1080ti".

Lol you're suggesting Nvidia is incompetent enough to release a new 1080ti at 4% faster than the 980ti?

That's actually just disrespectful to Nvidia.

 

[zentan]

The amount of time price/mm2 comes up without proper context is really beyond stupid.And the same arguments that are applied to gaming GPUs and CPUs for price/mm2 are somehow not applicable for Firepros/Quadros/Teslas ?
It's been pay for performance not /mm2.Nobody is forcing you to buy their products.
These are all companies dependent on profit for future survival and R&D,not a grocery shop.We have only 2 high performance GPU makers left as of now for reasons.Also implying AMD's margins are X so others must follow similar margins is stupid.They are the ones who are making losses and not others.If AMD gets Zen right they need to get as high margin as suitable in the market after all they have a debt to pay off which is not that far away and they are a leading edge company especially in graphics,would definitely need more money for R&D in future as various complexities go up.

 

[Det0x]

Lol you're suggesting Nvidia is incompetent enough to release a new 1080ti at 4% faster than the 980ti?

That's actually just disrespectful to Nvidia.

That would actually not surprise me, with these rumored specs

GP104 with half the die size with double precision added, compared to GM200..

How much denser are 16nm FinFET compared to 28nm ?

Anyhow, iam afraid this ain't the large performance jump many wants it to be. :'(

8b Transistors, 2560 CUDA Cores(?) and 256-bit bus VS 8b Transistors, 2816/3072 CUDA Cores and 384-bit bus

*Edit*

@ antihelten

huh? 20625 is only about 2% faster than a stock 980 Ti (20172), it's nowhere near the fastest aftermarket 980 Ti (~23000).

Unknown card:
Graphics score: 27271

GTX 980 Ti Waterforce:
Graphics score: 26112

Btw the numbers are from the same guy who made these videos

Okai so these numbers are fake then :thumbsdown:

 

[antihelten]

*Edit*

@ antihelten



Unknown card:
Graphics score: 27271

GTX 980 Ti Waterforce:
Graphics score: 26112

Yeah, I see that now, hadn't noticed 96Firebirds post before.

Anyway fake or not, I wouldn't be surprised if Nvidia's GP104 lineup was something like this:

1080 (non Ti) ~ aftermarket 980 Ti (20% faster than stock 980 Ti)
1070 ~ stock 980 Ti
1060 Ti ~ stock 980

Maybe 5% advantage for all 3 versus their predecessors, just so Nvidia can claim to be faster than last gen.

At that performance level Nvidia would probably not be able to get away with anything higher than $550 for the 1080, since the only real advantage versus the older GM200/GM204 chips would be a lower TDP.

Then down the line they can release a 400-450mm2 GP102 card as the 1080 Ti at the $650-700 price mark, and even further down the line they could finally release the full 600mm2 gaming focused GPU (i.e. GP100 without DP units). This would allow them to stretch out the 16nm node for a good long time, somewhat similar to what we saw with 28nm.

 

[MrTeal]

That would actually not surprise me, with these rumored specs

GK104 with half the die size with double precision added, compared to GM200..

How much denser are 16nm FinFET compared to 28nm ?

Anyhow, iam afraid this ain't the large performance jump many wants it to be. :'(

*Edit*

@ antihelten



Unknown card:
Graphics score: 27271

GTX 980 Ti Waterforce:
Graphics score: 26112



Okai so these numbers are fake then :thumbsdown:

Why would you think GP104 would have DP units added in? There will probably be some as there are in Maxwell, but there's no reason to think that it will be it will be a DP monster as GP100 is.

 

[Timmah!]

Your employeer pays you more than you need to survive, right?

If you have any money left besides food, transport and housing. I assume you demanded a pay cut right away?

Business is about profit and expansion. Not about charity and pro bono work. Just as your own job.

I am sure i am going to need a mortgage, when/if i am ever going to get my own living place/whatever else, cause my employer certainly wont increase my paycheck just because i am going to have bigger expenses than i have nowadays.

If thats how that works in Denmark, good for you.

 

[Det0x]

Why would you think GP104 would have DP units added in? There will probably be some as there are in Maxwell, but there's no reason to think that it will be it will be a DP monster as GP100 is.

I think it will have more dp enabled compared to Maxwell v2 because it will be used for Tesla/Quadro cards, i could very well be mistaken. But correct me if iam wrong, but these are the leaked specs for far ?

7.4-7.9b Transistors, 2560 CUDA Cores(?) and 256-bit bus for a ~290-300mm2 GP104

VS

8b Transistors, 2816/3072 CUDA Cores and 384-bit bus for for a 601mm2 GM200

Iam just not seeing a large performance jump compared to the current high end. (GM200) So i agree with antihelten numbers:

1080 (non Ti) ~ aftermarket 980 Ti (20% faster than stock 980 Ti)
1070 ~ stock 980 Ti
1060 Ti ~ stock 980

*edit* Typo

 

[MrTeal]

Correct me if iam wrong, but these are the leaked specs for far:

7.4-7.9b Transistors, 2560 CUDA Cores(?) and 256-bit bus for ~290-300nm GK104
VS
8b Transistors, 2816/3072 CUDA Cores and 384-bit bus for for a ~600nm GM200

Iam not seeing a large performance jump... So i agree with antihelten numbers:

Are you talking about GK104 or GP104?

I don't disagree that a 2560 core Pascal die wouldn't be a big improvement (if at all) over GM200, just that there's no reason to think that much of a 300mm² die would be devoted to DP.

 

[airfathaaaaa]

Why would you think GP104 would have DP units added in? There will probably be some as there are in Maxwell, but there's no reason to think that it will be it will be a DP monster as GP100 is.

having dp cores on a daily basis card is moot its a waste of space

 

[Pinstripe]

Correct me if I'm wrong, but isn't DP needed for heavy compute shader use?

 

[MrTeal]

having dp cores on a daily basis card is moot its a waste of space

nVidia is unlikely to remove all the DP units as it precludes testing DP code on a standard GPU. Even something like a GTX 750 can do some DP compute, it's just slow. More likely is that we see something with a really low DP:SP ratio, like the 1:32 used in Maxwell 2.

 

[Det0x]

Are you talking about GK104 or GP104?

I don't disagree that a 2560 core Pascal die wouldn't be a big improvement (if at all) over GM200, just that there's no reason to think that much of a 300mm² die would be devoted to DP.

I meant GP104 all along, just got my letters mixed up.

 

[Cloudfire777]

Geforce X70 / GTX 1070 @ 3DMark 11 (Chiphell)

www.3dmark.com/3dm11/11162175

Score: P20625
Graphics Score: 27271

www.chiphell.com/thread-1565171-1-1.html

fake fake fake fake fake

This is his first 3DMark11 run before he found out he messed up and blocked the software from reading his info

http://www.3dmark.com/3dm11/11162092

Its a high overclock on a GTX 980Ti

 

[exar333]

Nobody blamed a building. But someone wants companies to work for free, better yet, at a loss. Just so they can get what they feel entitled to.

This.

I wonder how many people here complaining about NV also own iPhones or Galaxy phones? Should those phones also decrease in price since new features are few and far between and the profit margins are 'too high'? They only cost $100 in materials, so is it criminal to sell it for more? A bunch of entitled-ists here!!

Anyway, back on topic.

 

[exar333]

Yeah, I see that now, hadn't noticed 96Firebirds post before.

Anyway fake or not, I wouldn't be surprised if Nvidia's GP104 lineup was something like this:

1080 (non Ti) ~ aftermarket 980 Ti (20% faster than stock 980 Ti)
1070 ~ stock 980 Ti
1060 Ti ~ stock 980

Maybe 5% advantage for all 3 versus their predecessors, just so Nvidia can claim to be faster than last gen.

At that performance level Nvidia would probably not be able to get away with anything higher than $550 for the 1080, since the only real advantage versus the older GM200/GM204 chips would be a lower TDP.

Then down the line they can release a 400-450mm2 GP102 card as the 1080 Ti at the $650-700 price mark, and even further down the line they could finally release the full 600mm2 gaming focused GPU (i.e. GP100 without DP units). This would allow them to stretch out the 16nm node for a good long time, somewhat similar to what we saw with 28nm.

I don't think NV can get away with a card much less than 40% faster than the Ti with that same name 1080Ti, etc. With how good the current 980Tis OC, if the new card is 'only' 20% faster, and OCs a little worse (less mature process, this is an unknown) it would be a wash for the new card. Total fail there for upgrades.

NV, in some ways, hurt themselves a little with how good of a OCer the 980Ti is. The next card not only needs to be appreciably faster, but significantly faster than a current factory-OC'd option. That's not an easy job...

 

[poofyhairguy]

I wonder how many people here complaining about NV also own iPhones or Galaxy phones? Should those phones also decrease in price since new features are few and far between and the profit margins are 'too high'?

Uh, yes.

I will probably never again own another $700 flagship phone. They aren't worth it.

And Samsung/Apple knows that. The S7 came out cheaper than any recent Galaxy and Apple released a very high value iPhone SE.

 

[antihelten]

I don't think NV can get away with a card much less than 40% faster than the Ti with that same name 1080Ti, etc. With how good the current 980Tis OC, if the new card is 'only' 20% faster, and OCs a little worse (less mature process, this is an unknown) it would be a wash for the new card. Total fail there for upgrades.

NV, in some ways, hurt themselves a little with how good of a OCer the 980Ti is. The next card not only needs to be appreciably faster, but significantly faster than a current factory-OC'd option. That's not an easy job...

The kind of upgrade I described (i.e. only a very small improvement performance wise , slight drop in price and significant improvement in TDP), is basically what we saw with the release of 980 and 970, and they were both hugely successful upgrades from Nvidia's point of view (i.e. sales).

With regards to the overclockability of GM200, I agree that it was probably a bit more than what Nvidia would have prefered, and it would certainly have been easier for them to deliver a compelling product if it hadn't been so good.

 

[exar333]

The kind of upgrade I described (i.e. only a very small improvement performance wise , slight drop in price and significant improvement in TDP), is basically what we saw with the release of 980 and 970, and they were both hugely successful upgrades from Nvidia's point of view (i.e. sales).

With regards to the overclockability of GM200, I agree that it was probably a bit more than what Nvidia would have prefered, and it would certainly have been easier for them to deliver a compelling product if it hadn't been so good.

While I don't disagree, the move to 980 was on the same node, on a mature process, and rather incremental. Something similar here would be almost a total fail, unless they positioned it as a true 'mid-range' product and not a Ti model. The name matters.

 

[Kris194]

1080 (non Ti) ~ aftermarket 980 Ti (20% faster than stock 980 Ti)
1070 ~ stock 980 Ti
1060 Ti ~ stock 980

No way that rumored 2560CC will be barely 20% faster than stock 980 TI, it would be ridiculous and it would mean that it is only 10% faster than Titan X. If full GP104 have 2560CC it would be like

~30 - 40% faster clocks than Titan X
~83% units of Titan X
At least 10% more efficient SM what would be equal to 2816 CUDA Cores (GTX 980 Ti has exactly that amount and look what ~1,3GHz GTX 980 Ti is doing with stock Titan X

http://pclab.pl/zdjecia/artykuly/chaostheory/2016/01/gfx/OC/oc_bf4_2560u.png

 

[antihelten]

No way that rumored 2560CC will be barely 20% faster than 980 TI, it would be ridiculous.

980 Ti is 2816CC or 10% more than 2560, so for a 2560CC GP104 to be 20% faster would require it getting 32% more performance from IPC increases and/or frequency increases.

If GP104 boosts to 1480 like GP100, that alone would be roughly 25% more than a stock 980 Ti (which boosts to about 1200 on average), so it would then need 5-10% more IPC to end up at 20% fast in total than a stock 980 Ti. So far there has been no indications as far as I can tell that Pascal is bringing any IPC improvements as far as gaming is concerned (all the news have been about compute), so I would say it is prudent not to expect too much here, at this point in time.

Also 20% faster than 980 Ti would be about 17% faster than Titan X, not 10%. There really isn't much of a gap between those two.

Edit:
If we compare to Titan X instead of 980 Ti we get much the same. Titan X boosts to about 1125 on average, so a 1480 GP104 would be ~32% higher, not 40%. So you have 32% higher clock, 17% fewer cores, and 10% higher IPC: 1.32*0.83*1.1 = 1.21 or 21% faster than a Titan X, which would be the same as ~23% faster than 980 Ti.

While I don't disagree, the move to 980 was on the same node, on a mature process, and rather incremental. Something similar here would be almost a total fail, unless they positioned it as a true 'mid-range' product and not a Ti model. The name matters.

I very much doubt the average buyer judges the value of a GPU based upon which node it is manufactured, so I doubt it would have much of an effect on Nvidia's sales.

Sure it would probably be quite poorly received on forums like this one, but we do not make up the bulk of customers.

Also I'm assuming here that the top end GP104 would be named 1080 and not 1080 Ti, since a 1080 Ti performing similarly to an aftermarket 980 Ti would look quite bad.

 

[3DVagabond]

I agree that price should be based on performance and not mm2.

Performance relative to what? Ideally price would be based on cost. It's always based on perceived market value, though.

 

[antihelten]

Performance relative to what? Ideally price would be based on cost. It's always based on perceived market value, though.

Ideally price would be based on value not cost, otherwise there's no incentive for companies to reduce their costs.

 

[Kris194]

antihelten, notice that you don't know real gpu boost clock for GP100. Nvidia says that GTX 980 Ti boosts to 1075 MHz but like you noticed, many of them boosts to ~1200 MHz

 

[antihelten]

antihelten, notice that you don't know real gpu boost clock for GP100. Nvidia says that GTX 980 Ti boosts to 1075 MHz but like you noticed, many of them boosts to ~1200 MHz

Of course I don't know, but until we have more information the safest bet is probably to assume that it's not any higher than what has been announced for GP100.