[ Various] NVIDIA could launch Pascal-based mobile GPUs by end-2016

[maddie]

I don't think GP100 is anywhere near 500 mm2. I am thinking in the 350-400 range. Even then, they'll just have to deal and be creative in binning... plus charging $1500 for a Titan and $5K+ for Teslas (2 die) will help for sure. GP104 is probably close to Polaris 11's size, maybe a bit bigger. GP102 is probably in the 300 range.

Yes, I know. You should downgrade your performance expectations for this generation pretty significantly unless you are willing to pay.

Did I miss some information? Is the size of Polaris 11 known?

 

[jpiniero]

Did I miss some information? Is the size of Polaris 11 known?

No... but if Polaris 10 is in the 125 mm2 range, it'd make sense if it was in the 175-200 mm2 range. Performance comparable to Grenada.

 

[JDG1980]

I don't think GP100 is anywhere near 500 mm2. I am thinking in the 350-400 range.

That would be completely unprecedented. Here's a list of the die sizes of each of Nvidia's '0'-series GPUs, going back to Tesla:
GT200 - 576mm^2 (65nm)
GT200B - 470mm^2 (55nm)
GF100 - 529mm^2 (40nm)
GF110 - 520mm^2 (40nm)
GK110 - 551mm^2 (28nm)
GK210 - 561mm^2 (28nm)
GM200 - 600mm^2 (28nm)

Only one entry below 500mm^2, and that's just because it was a half-node shrink of an existing chip rather than a new design. I can see them going down to 500mm^2 for 16nm FF+, but no smaller. This is the big HPC chip; it has to beat Xeon Phi. They can't afford to half-ass it.

Credible rumors say that GP100 has almost 17 billion transistors - there's no way that would possibly fit on 350-400mm^2. That's actually more than double the transistor count (8 billion) of GM200.

 

[JDG1980]

No... but if Polaris 10 is in the 125 mm2 range, it'd make sense if it was in the 175-200 mm2 range. Performance comparable to Grenada.

I'd expect a bigger gap than that, considering that AMD is only going to have two FinFET GPUs this year to cover their lineup. Why would they position them so close together? In first-gen 28nm we saw Cape Verde at 123mm^2 and Pitcairn at 212mm^2. We're pretty much all expecting Polaris 10 to be about the same die size as Cape Verde, so I would be surprised if Polaris 11 is any smaller than Pitcairn. I agree that the performance of Polaris 11 will probably land around what we currently get from Hawaii/Grenada, maybe a little better due to architectural improvements.

As for Nvidia's GP104, remember that a '4'-series chip has to be able to beat the last generation's '0'-series chip, and by a decent margin. Thus, GP104 has to surpass GM200 (GTX 980 Ti / Titan X), and not just by a hair. That is going to require a chip no smaller than 325mm^2. Even this is stretching it, since GM204 is 398mm^2. The memory controllers can save some space by staying at 256-bit and going to GDDR5X for more bandwidth, but doubling the shader count to 4096 and ROP count to 128 is going to need lots of transistors. And that's what is going to have to happen to beat GM200, since by most accounts, Pascal won't see too many architectural improvements over Maxwell, and those it does see will mostly be compute-focused.

 

[Vesku]

Doubt Polaris 11 is much less than 300mm2, AMD only has 10 and 11 for 2016 and they said they will have a mix of GDDR5 low end and HBM2 for the upper end.

Looking like it will be Polaris 11 HBM2 first out the gate in terms of "performance" gaming GPU with Nvidia GP104 GDDR5X showing up a bit later.

 

[3DVagabond]

I saw a post that showed 3 new GPU's in AMD's drivers. I can't find it now, but will continue trying. There are rumors, despite what Raja is saying that there are 3 "Polaris" GPU's in house now.

 

[littleg]

I saw a post that showed 3 new GPU's in AMD's drivers. I can't find it now, but will continue trying. There are rumors, despite what Raja is saying that there are 3 "Polaris" GPU's in house now.

Thing is AMD is a publically traded company. If he comes out and says they have two GPUs when they really have three then he's going to be in big trouble.

 

[JDG1980]

Doubt Polaris 11 is much less than 300mm2, AMD only has 10 and 11 for 2016 and they said they will have a mix of GDDR5 low end and HBM2 for the upper end.

Here's what AMD actually said, according to Anandtech's article:

First and foremost, Polaris will encompass both GDDR5 and High Bandwidth Memory (HBM) products. Where the line will be drawn has not been disclosed, but keeping in mind that HBM is still a newer technology, it’s reasonable to expect that we’ll only see HBM on higher-end parts. Meanwhile the rest of the Polaris lineup will continue to use GDDR5, something that is not surprising given the lesser bandwidth needs of lower-end parts and the greater cost sensitivity.


Meanwhile RTG has also disclosed that the first Polaris parts are GDDR5 based. Going hand-in-hand with what I mentioned earlier about RTG’s Polaris demonstration, it seems likely that this means we’ll see the lower-end Polaris parts first, with high-end parts to follow.

Note that if you read carefully, all that was volunteered is that some high-end Polaris parts will use HBM. The first Polaris parts (plural) will use GDDR5. Nowhere was it specified that either of the two Polaris chips coming in 2016 will be among those with HBM.


It's still possible that AMD could go for the crown with a >300mm^2 part. But it's equally possible that they will try to sweep the laptop market (and <75W desktop) and midrange by putting out two smaller chips first. That would also have the advantage of higher yields and a more competitive price point.

 

[Pariah]

Thing is AMD is a publically traded company. If he comes out and says they have two GPUs when they really have three then he's going to be in big trouble.

No, he won't. Keeping products a secret from the public does not get executives in big trouble. In this case, if there actually are 3 chips ready to go this year, it doesn't make much sense to say there are only 2. For that reason alone, it's unlikely there are 3. Just having a driver place holder doesn't mean a release is imminent, or even guaranteed.

 

[3DVagabond]

Thing is AMD is a publically traded company. If he comes out and says they have two GPUs when they really have three then he's going to be in big trouble.

They claimed no new GPU's in 2013 and then they released Hawaii that year. There's no law that states a company has to publicly disclose future product releases.

No, he won't. Keeping products a secret from the public does not get executives in big trouble. In this case, if there actually are 3 chips ready to go this year, it doesn't make much sense to say there are only 2. For that reason alone, it's unlikely there are 3. Just having a driver place holder doesn't mean a release is imminent, or even guaranteed.

Never said imminent or anything was or wasn't going to be released.

 

[Pariah]

They claimed no new GPU's in 2013 and then they released Hawaii that year. There's no law that states a company has to publicly disclose future product releases.

No idea what AMD was trying to do with that announcement, but they only stuck to that story for a week and then (Feb 15th):

http://www.pcworld.com/article/2028408/amd-to-release-new-radeon-hd-8000-graphics-cards-in-2013.html

"It's official. Bucking a recent round of rumors, AMD revealed today that it will be releasing new Radeon desktop GPUs "from top to bottom" in 2013, though you won't be seeing the next-gen Radeon 8000 series graphics cards until late in the year."

Never said imminent or anything was or wasn't going to be released.

So you said nothing? I thought I saw something there, I guess not...



Edit: Bolding added to important text 3DVagabond missed the first time about 8000 series release date in 2013.

 

[3DVagabond]

No idea what AMD was trying to do with that announcement, but they only stuck to that story for a week and then (Feb 15th):

http://www.pcworld.com/article/2028408/amd-to-release-new-radeon-hd-8000-graphics-cards-in-2013.html

"It's official. Bucking a recent round of rumors, AMD revealed today that it will be releasing new Radeon desktop GPUs "from top to bottom" in 2013, though you won't be seeing the next-gen Radeon 8000 series graphics cards until late in the year."



So you said nothing? I thought I saw something there, I guess not...

If you read the article you'll see that it says you'll only see 7000 series. No 8000 series, which was what people were calling AMD's next gen at the time.

Between that and then your smart comment at the end, I'd say your reading comprehension is why you didn't understand.

 

[Pariah]

If you read the article you'll see that it says you'll only see 7000 series. No 8000 series, which was what people were calling AMD's next gen at the time.

I don't really know what to say here. I guess I'll let someone else try to explain to you what the article says about the 8000 series release date. It was in the 1st paragraph, I quoted it for you in my post, and you still missed it. There are no words...

One other thing you missed, the title of the article:

"AMD to release new Radeon HD 8000 graphics cards in 2013"

The irony of your reading comprehension line is fantastic.

 

[AtenRa]

Just to point out that if Polaris 10 is ~125mm2 it will be close to Tonga (R9 380/X) and GM206 (GTX960) performance or higher. Perhaps even better than those two 28nm at DX-12 games.

 

[Vesku]

Just to point out that if Polaris 10 is ~125mm2 it will be close to Tonga (R9 380/X) and GM206 (GTX960) performance or higher. Perhaps even better than those two 28nm at DX-12 games.

A straight die shrink would put the GTX 960 close to the same size. I don't see AMD pulling too far ahead in raw performance if at all. They seem quite confident about Perf/W, though. Will be interesting to see if GPU reviewers put as much weight behind latency and perf/W as they have with 28nm.

 

[3DVagabond]

I don't really know what to say here. I guess I'll let someone else try to explain to you what the article says about the 8000 series release date. It was in the 1st paragraph, I quoted it for you in my post, and you still missed it. There are no words...

One other thing you missed, the title of the article:

"AMD to release new Radeon HD 8000 graphics cards in 2013"

The irony of your reading comprehension line is fantastic.

You are right. I actually misread. Interesting that they didn't do what the article says though. They only released one new card, Hawaii IIRC. They did release an 8000 series, but it was simply rebrands for OEM's. Still, reading this only shows that you can take what is said with a grain of salt.

A straight die shrink would put the GTX 960 close to the same size. I don't see AMD pulling too far ahead in raw performance if at all. They seem quite confident about Perf/W, though. Will be interesting to see if GPU reviewers put as much weight behind latency and perf/W as they have with 28nm.

I have no doubt that if both companies execute together that nVidia will come up with a good GPU. With no Pascal to show yet and AMD has shown Polaris, and nVidia's history of die shrink and new memory tech issues, I'll be surprised if AMD doesn't release earlier. Some people are saying 2 qtrs. earlier, but that remains to be seen.

 

[sontin]

His comparision doesnt make any sense. Tonga is 360mm^2. Nearly the same size like GM204.

So why would only AMD able to reduce the die size ~3x but not nVidia? :\

 

[AtenRa]

His comparision doesnt make any sense. Tonga is 360mm^2. Nearly the same size like GM204.

So why would only AMD able to reduce the die size ~3x but not nVidia? :\

Tonga has a 384bit memory controller and GM206 only 128bit memory controller.

Also Tonga is way faster than GM206, but i was only talking about performance not die sizes.

 

[sontin]

Tonga is way faster because it is way bigger and uses twice the power than GM206. It doesnt make sense to compare Tonga with GM206 from a performance standpoint alone.

A 120mm^2 Polaris will not be fast as GM206. That would mean that AMD is able to improve performance around ~3 from a size perspective.

 

[AtenRa]

Tonga is way faster because it is way bigger and uses twice the power than GM206. It doesnt make sense to compare Tonga with GM206 from a performance standpoint alone.

A 120mm^2 Polaris will not be fast as GM206. That would mean that AMD is able to improve performance around ~3 from a size perspective.

Take off 2x64 bit memory controllers and 4 CUs from Tonga and you have a 300mm2 die. At 14nm it should be close to 120-130mm2.

Now add all the enhancements made to Polaris over Tonga/Fury and we could have a 125mm2 die at 14nm with R9 380X performance at half the power.

 

[sontin]

No, the die size would not be so much smaller. Look at Hawaii with 40% more CUs and 33% more memory controller. It is just 22% bigger or 80mm^2

Even using your numbers: nVidia could bring GM206 down from 220mm^2 to 100mm^2 with a power consumption of 44W.

Looks much better than 120mm^2 and 75W, or?

 

[PPB]

Tonga is way faster because it is way bigger and uses twice the power than GM206. It doesnt make sense to compare Tonga with GM206 from a performance standpoint alone.

A 120mm^2 Polaris will not be fast as GM206. That would mean that AMD is able to improve performance around ~3 from a size perspective.

It will be, as 270x now is tied to the GM206 according to TP's averages. People do a misconception and use 380x/Tonga as comparison when the GPU that is now dead even with GM206 is Pitcarin (an early 2012 design that only +20% power consumption at same performance and 5% less die area compared to GM206).

It also makes sense that it will be a 125mm2 die, scaling is 0.55x aprox. in 14LPP compared to 28SHP. Add some IPC to the shaders and you have your Polaris 10. Also 2.5 better perf/watt compared to Pitcarin puts it right at 50w.

People argue why Pitcarin go recycled so much all these years, it was their best perf/watt design considering how well the performance improved all these years. Tahiti/Tonga were always overbuilt, both in transistor qty. (Tahiti had a lot of redundancy to achieve desirable yields and beat GK104 to market) and bus width. Tonga's 386b bus doesnt make any sense considering it has better effective bandwith compared to Tahiti. Considering GDDR5X is out, AMD should put more stress on RAM speeds and less on bus width on non-HBM designs (but not cripple the xbar/imc to 970 levels either :awe

 

[AtenRa]

No, the die size would not be so much smaller. Look at Hawaii with 40% more CUs and 33% more memory controller. It is just 22% bigger or 80mm^2

Even using your numbers: nVidia could bring GM206 down from 220mm^2 to 100mm^2 with a power consumption of 44W.

Looks much better than 120mm^2 and 75W, or?

For the second time in a few minutes, I say again im not comparing Polaris to whatever NVIDIA will bring at 16nm. Im comparing a 125mm2 Polaris at 14nm vs Tonga.

 

[jpiniero]

Credible rumors say that GP100 has almost 17 billion transistors - there's no way that would possibly fit on 350-400mm^2. That's actually more than double the transistor count (8 billion) of GM200.

17B does kind of feel that it would be too much to get into 400 mm2 but at the same time (as you mentioned) 500 mm2 feels like it'd be way too big for any kind of yield. So we'll have to see. But they are going to need to release a Titan to help with the volume, that part is for sure.

 

[sontin]

GM206 is much faster than Pitcarn: http://www.hardware.fr/articles/945-6/consommation-efficacite-energetique.html

And GM206 is much more advanced. Pitcairn is missing a lot of features which would make the die much bigger.