NVIDIA Pascal Thread

[Head1985]

I think Nvidia will go for the higher prices as they need to sustain those record margins. Yields on 16FF+ are much lower than 28nm (especially for chips much larger than mobile SoCs which are tiny at 100 sq mm) and wafer cost is much higher.

I really dont care.What i care is performance and price.If they cant do it then i will not buy them simple as that.I dont care if they want record margins and profits.

Also remember 28nm WAS new when GTX680/670/660TI hit.And cost fortune and has bad yields.

 

[Silverforce11]

And this is why people shouldn't hate on AMD. Having AMD being competitive is a good thing (unless they illegally collude!)...

So you guys should hope Polaris is going to be awesome and cheap.

 

[Head1985]

680 wasn't 35% faster than a 580 (only about 25%), and the 580 didn't cost $650 like 980 Ti.
670 was $400 as you said, but the 570 was only $350, so a price hike of $50 or ~15%. So the situation isn't really comparable.

If you want to draw parallels to the 500 to 600 series transition, then this is what we would see:

1080: 980 + 25-30% (980 Ti + 0-5%) - $500
1070: 980 +20% (980 Ti -5%) - $400
1060 Ti: 980 +5% (980 Ti -15%) - $300

You know GTX980 is mainstream SKu.You really dont know what are you talking about.
GTX580 was todays TITANX.
So when GTX680 was 30% faster than GTX580, then 1080 should be 30% faster than TITANX.

To be clear its like this:
GTX560TI-GTX680-GTX980-GTX1080
GTX580-GTX780TI-GTXTITANX-Big pascal

 

[Timmah!]

I think Nvidia will go for the higher prices as they need to sustain those record margins. Yields on 16FF+ are much lower than 28nm (especially for chips much larger than mobile SoCs which are tiny at 100 sq mm) and wafer cost is much higher. So Nvidia will price it up and see how AMD responds. If Polaris 10 gets too close to GTX 1070 and comes in at USD 350-400 then it might force Nvidia to respond otherwise they will just milk the market (which they have mastered).

Any actual proof or at least reference that these 300mm2 GP104s yield badly?

BTW, if GP104 launch is sort of meant to be closest in similarity to GK104 (just about 25 percent above top chip of previous generation), what do you think are the chances we will see dual GP104 board this year, to compete with Radeon Pro Duo for performance crown? Something akin GTX690?

 

[Timmah!]

If you want to draw parallels to the 500 to 600 series transition, then this is what we would see:

1080: 980 + 25-30% (980 Ti + 0-5%) - $500
1070: 980 +20% (980 Ti -5%) - $400
1060 Ti: 980 +5% (980 Ti -15%) - $300

.

No, because the 580 was fully unlocked top chip of its generation. Something 980 is not, Titan X is. Just because in case of Titan they doubled the price in effort to turn it into halo product, it does not make it more than what 580 once was.

 

[Glo.]

If it's really 320 mm2 then the $499 1070 rumors are probably true.



It looks like Pascal's efficiency has gotten "worse" though... enough that AMD has caught up with Polaris. The 1070 and 1080 should be faster because of the higher clock speeds but at higher TDP.

I would not say worse.

GP104-200 - 2048 CUDA core pascal GPU with 1480 MHz boost clock will be most probably 165W.
GP104-400 - 2560 CUDA core GPU with 1480 MHz boost clock will likely be slightly under 200W.

That is because process brings 60% lower power consumption or 30% higher speed, at the same thermal envelope.

So the words that GTX 1070 with 30% higher performance from GTX 980 might be true.
I tend to expect from GTX 1080 25-30% higher performance from GTX980 Ti.

Which would fall in line perfectly for 649$ price bracket.

How will Polaris compare to that? Depends on... everything to be honest. 3072 GCN4 GPU with 1150 MHz will mop the floor with GTX 1080, while using much less power. 2560 GCN4 chip might be within 5%(slower/faster) than 2560 CUDA core GPU while using much less fuel.

 

[antihelten]

You know GTX980 is mainstream SKu.You really dont know what are you talking about.
GTX580 was todays TITANX.
So when GTX680 was 30% faster than GTX580, then 1080 should be 30% faster than TITANX.

To be clear its like this:
GTX560TI-GTX680-GTX980-GTX1080
GTX580-GTX780TI-GTXTITANX-Big pascal

The GTX 580 was certainly not today's Titan X, The GF110 chip was basically a slightly fixed GF100 chip, which means that the closest comparison would be the GK110 chip aka Titan non-X (which is basically the same speed as the GTX 980).

Nvidia has never before had an intranode upgrade that changed the architecture as fundamentally as Maxwell did nor has Nvidia ever before produced a chip as big as GM200 (15% bigger than GF110). So there is really no precedent for what the GM200 chip should or should not cost, the closest chip size wise is the GT200 chip at 576mm2, guess what that cost?

So be clear it is not:

GTX560TI-GTX680-GTX980-GTX1080
GTX580-GTX780TI-GTXTITANX-Big pascal

Or whatever the hell you may imagine. It is whatever the hell Nvidia wants it to be, and last I checked Nvidia was run by Jen-Hsun Huang, not Head1985.

So if you don't mind I'll continue basing my prediction based upon Nvidia's history, not your dreams and hopes.

 

[Glo.]

Second thing is that those chips are qualification samples without even metal cover in the die.

They are also dated for 14 Week of 2016. Which means they were produced around 3 weeks ago. So It looks like 3 months for release, at least GP104-400 model with GDDR5X, which would fall in line perfectly with July launch for that chip.

 

[Sweepr]

A 256-bit GDDR5 VGA, described by AMD itself as 'desktop mainstream' matching or mopping the floor with NVIDIA's new rumoured $649 Geforce GTX 980 Ti replacement card (and Fury X as well). Not going to throw around performance predictions here but I think some people are being a tad overly optimistic, as usual.

 

[Glo.]

A 256-bit GDDR5 VGA, described by AMD itself as 'desktop mainstream' matching or mopping the floor with NVIDIA's new rumoured $649 Geforce GTX 980 Ti replacement card (and Fury X as well). Not going to throw around performance predictions here but I think some people are being a tad overly optimistic, as usual.

Maybe I have overextended it a bit, but it will be faster, because SIMD vs SIMD performance will be extremely similar, but AMD will have few features that will help with Gameworks titles(Primitive Discard Accelerator is simply Anti-Gameworks feature), Async performance, and new Graphics Schedulers.

649$ GTX 980 TI replacement will also have 256 Bit memory bus, and we still don't know if AMD GPUs will not in the end use GDDR5X. Maybe, maybe not.

Yes I tend to be optimistic. However I also have cool mind about this.

 

[raghu78]

Any actual proof or at least reference that these 300mm2 GP104s yield badly?

BTW, if GP104 launch is sort of meant to be closest in similarity to GK104 (just about 25 percent above top chip of previous generation), what do you think are the chances we will see dual GP104 board this year, to compete with Radeon Pro Duo for performance crown? Something akin GTX690?

Yields for mobile SoC chips like Kirin 950 were 80% in Q3 2016.

http://www.anandtech.com/show/9762/hisilicon-announces-kirin-950-huawei

"In fact, HiSilicon explains that along with Apple they've been the two main lead partners of the Taiwanese semiconductor giant, and both parties have been working closely together to try to improve the design and to tune the process. In fact, the company revealed that first mass production (also commonly named as risk production) started as early as last January. Over the following months both companies cooperated to sort out bugs and imperfections in the design (chip revisions) to go up from 20% yield in the earliest runs to up to 80% yields and qualified mass production this last August."

Those chips are tiny at <= 100 sq mm. With increasing die size those yields decrease in a non-linear fashion ie yield rate decreases more rapidly. Other than TSMC and their customers nobody will know the actual yield rates of dies > 300 sq mm. I can speculate that GM200 at 28nm is yielding better than GP104 at 16FF+ as its an extremely mature node as its 4.5 years old while 16FF+ is 9 months old . If you add in the significantly higher wafer costs (28nm is dirt cheap at roughly half the wafer costs of 16FF+) I think Nvidia can produce GTX 980 Ti for cheaper than the flagship GTX 1080 as it also uses GDDR5X which is costlier and just beginning to ramp (so lower yields than GDDR5). Anyway I think Nvidia's prices will fall only if AMD brings something extremely compelling in perf/$ which forces Nvidia to revisit their pricing otherwise Nvidia will be happy to sustain those record margins.

 

[Sweepr]

Pascal / Polaris prices relative to Geforce GTX 980 Ti according to Chiphell user gtx9:

1080 = 135% 5500RMB level
1070 = 110% 4000RMB grade
980 Ti = 100%
Desktop Polaris P10 = 80% 3000RMB level

1 USD = ~6.5 RMB

Ps: Assuming there are 3 GP104 VGAs, maybe they got the naming wrong. I read somewhere there could be two Geforce GTX 1080 versions, really doubt Geforce GTX 1070 will cost this much.

 

[Paul98]

What changes have been made to increase game performance? I see a lot of stuff for HPC, not much for gaming.

 

[antihelten]

Pascal / Polaris prices relative to Geforce GTX 980 Ti according to Chiphell user gtx9:



1 USD = ~6.5 RMB

Ps: Assuming there are 3 GP104 VGAs, maybe they got the naming wrong. I read somewhere there could be two Geforce GTX 1080 versions, really doubt Geforce GTX 1070 will cost this much.

For reference the 980 Ti also sold for ~5500 RMB in China, so those prices would indicate:

1080 - $650
1070 - $500

A theoretical 1060 Ti would then most likely be $400.

I really hope those prices are only speculation though, since otherwise Polaris looks kinda crappy. Basically an R9 390 +5-10% for $350. Of course this also goes completely against AMD's promises of providing 390 level performance at a lower price point.

 

[airfathaaaaa]

For reference the 980 Ti also sold for ~5500 RMB in China, so those prices would indicate:

1080 - $650
1070 - $500

A theoretical 1060 Ti would then most likely be $400.

but if a freaking 1080(not even a ti) will cost 5500 aka todays 850 dollars almost
what will cost a ti? 1000+?

 

[antihelten]

but if a freaking 1080(not even a ti) will cost 5500 aka todays 850 dollars almost
what will cost a ti? 1000+?

Again a 980 Ti also costs 5500 RMB, but it doesn't cost $850 in the US. You can't just do a straight currency conversion.

980 Ti is currently 18% more expensive than the 980 ($100), so I guess if Nvidia releases a 1080 Ti on top of the cards from the ChipHell post, then it would also be 18% or $100 higher, which would put it at ~$750.

Nvidia has previously shown that they are willing to push the X80 Ti level GPU above $650 when they have little to no competition, seeing as the 780 Ti was $700, so I guess it wouldn't be super surprising to see a 1080 Ti in that range, but it would depend upon what AMD can get out (Vega).

 

[airfathaaaaa]

Again a 980 Ti also costs 5500 RMB, but it doesn't cost $850 in the US. You can't just do a straight currency conversion.

980 Ti is currently 18% more expensive than the 980 ($100), so I guess if Nvidia releases a 1080 Ti on top of the cards from the ChipHell post, then it would also be 18% or $100 higher, which would put it at ~$750.

Nvidia has previously shown that they are willing to push the X80 Ti level GPU above $650 when they have little to no competition, seeing as the 780 Ti was $700, so I guess it wouldn't be super surprising to see a 1080 Ti in that range, but it would depend upon what AMD can get out (Vega).

but im not talking about us im talking about china 850+ for a card that isnt even the best version of it is just hmmm wtf
(agree doesnt make much sense talking about their currency but still its a BIG difference compared the situation on china)

 

[antihelten]

but im not talking about us im talking about china 850+ for a card that isnt even the best version of it is just hmmm wtf
(agree doesnt make much sense talking about their currency but still its a BIG difference compared the situation on china)

Sure, but unless you live in China, the actual prices themselves are uninteresting, the interesting thing is the relative prices in China, since those should be similar to what you find in the rest of the world (e.g. if a 980 Ti is 18% more expensive than a 980 in the US, then it should also be 18% more expensive in China).

So what the ChipHell numbers tell us is this:

Price:
1080 = 980 Ti
1070 = 86% of 980 (this would equate to roughly $470, but that's a slightly odd number so most likely it will be $500 in the US)

 

[AtenRa]

If they will price the GP104 at $650+ it will be a new record of silliness. I sure hope they will not do something like that and price it at $550 tops for aftermarket cards.

 

[Timmah!]

Yields for mobile SoC chips like Kirin 950 were 80% in Q3 2016.

http://www.anandtech.com/show/9762/hisilicon-announces-kirin-950-huawei

"In fact, HiSilicon explains that along with Apple they've been the two main lead partners of the Taiwanese semiconductor giant, and both parties have been working closely together to try to improve the design and to tune the process. In fact, the company revealed that first mass production (also commonly named as risk production) started as early as last January. Over the following months both companies cooperated to sort out bugs and imperfections in the design (chip revisions) to go up from 20% yield in the earliest runs to up to 80% yields and qualified mass production this last August."

So they improved their yields from 20 to 80 percent between January and August last year. Now its almost May, more than half year later. Obviously 300 mm2 is bigger than 100 mm2, but going by that improvement rate, dont you think that automatically assuming yields are terrible is bit of an doom-mongering? Surely the experience TSMC gets by that process tuning done back then helps them later when producing other bigger chips?

 

[AtenRa]

I dont believe yields are terrible but im sure 28nm yields of 300mm2 chips are higher today than 16nm.

 

[raghu78]

So they improved their yields from 20 to 80 percent between January and August last year. Now its almost May, more than half year later. Obviously 300 mm2 is bigger than 100 mm2, but going by that improvement rate, dont you think that automatically assuming yields are terrible is bit of an doom-mongering? Surely the experience TSMC gets by that process tuning done back then helps them later when producing other bigger chips?

You have to remember TSMC did not start volume production till late Q2/early Q3 2016. So the steep yield rate improvement from 20 to 80% during Jan to August 2016 was happening during most of risk production which is how it happens. Expecting the same rate of improvement during volume production ramp is not realistic. btw the yield rate on 100 sq mm chips does not mean similar yields of > 300 sq mm chips as yield rate decreases non linearly with increase in die size.

 

[airfathaaaaa]

You have to remember TSMC did not start volume production till late Q2/early Q3 2016. So the steep yield rate improvement from 20 to 80% during Jan to August 2016 was happening during most of risk production which is how it happens. Expecting the same rate of improvement during volume production ramp is not realistic. btw the yield rate on 100 sq mm chips does not mean similar yields of > 300 sq mm chips as yield rate decreases non linearly with increase in die size.

according to tsmc they had "mature" node fabrication a year ago

the same time that qualcomm gave them the middle finger and signed with samsung

 

[Adored]

So they improved their yields from 20 to 80 percent between January and August last year. Now its almost May, more than half year later. Obviously 300 mm2 is bigger than 100 mm2, but going by that improvement rate, dont you think that automatically assuming yields are terrible is bit of an doom-mongering? Surely the experience TSMC gets by that process tuning done back then helps them later when producing other bigger chips?

If we assume 80% yield on a 100mm2 chip, we're talking around 55% yield for a 316mm2 GPU with the same defect density as Samsung claimed at 0.2/cm.

So definitely not terrible, but also not amazing. It'll likely be less as there is more to yield than just defects, so under 50% I'd say.

 

[ShintaiDK]

A 256-bit GDDR5 VGA, described by AMD itself as 'desktop mainstream' matching or mopping the floor with NVIDIA's new rumoured $649 Geforce GTX 980 Ti replacement card (and Fury X as well). Not going to throw around performance predictions here but I think some people are being a tad overly optimistic, as usual.

Understatement of the year