NVIDIA Pascal Thread

[Erenhardt]

GP104 will be much, much faster than Polaris 10 if Polaris 10 is around 240mm^2.

Very convincing. We will hopefully know tomorrow how good pascal is.

 

[Glo.]

GP104 will be much, much faster than Polaris 10 if Polaris 10 is around 240mm^2.

May I know the reasons behind this?

 

[sontin]

May I know the reasons behind this?

GM200 has 8 billion transistors, Fij has 9,2 billions and is only as fast as GM200.

GP104 should be around 8 billions (25 million transistors per mm^2) while Polaris 10 should be around 7 billions (~30 million per mm^2).

There doesnt exist any reason to believe that AMD will beat nVidia on performance per transistor with the 16nm generation.

 

[Glo.]

One word. Hardware. Think a second. Which GPU will be faster? 232mm2 3072 GCN core GPU with 1150 MHz with 125W of TDP, or 330mm2 2560 CUDA core GPU with 1.48 GHz turbo core. with around 200W TDP?

No, even in compute terms of performance Pascal will be faster. 7 TFLOPs, vs 7.6 TFLOPs of performance. But overall both will be extremely close. Add to that possibility that Polaris 10 will be 300-350$ GPU and GP104 might even be 650$. You get the picture in what situation we can be in future?

I do not buy s*** that most people sell in those threads, both from optimists, skeptics, and people who spread FUD, about magical, miraculous Nvidia hardware or AMD.

 

[xpea]

One word. Hardware. Think a second. Which GPU will be faster? 232mm2 3072 GCN core GPU with 1150 MHz with 125W of TDP, or 330mm2 2560 CUDA core GPU with 1.48 GHz turbo core. with around 200W TDP?

wow so optimistic about AMD and pessimistic about Nvidia

what about something more unbiased ?

1/ I highly doubt that GP104 will have higher TDP than GM204. Nvidia won't make this mistake. It should will be 150~180W for the top SKU with GDDR5X

2/ I highly doubt that Polaris 10 will have 3072 GCN cores. All leaks go for 2560, much more reasonable if 232mm2 is right

If we take into account 1+2, it will change completely the equation and it will be much closer to the reality...

 

[Riek]

GM200 has 8 billion transistors, Fij has 9,2 billions and is only as fast as GM200.

GP104 should be around 8 billions (25 million transistors per mm^2) while Polaris 10 should be around 7 billions (~30 million per mm^2).

There doesnt exist any reason to believe that AMD will beat nVidia on performance per transistor with the 16nm generation.

Hard to compare between older AMD generation and assume the relation is still the same. escpecially given the underwhelming performance between the specs of Fury X versus Hawai..

From the rumours we have:

Rumour 1: polaris 10 ~ 980ti
Rumour 2: 1060 ~970
Rumour 3: 1070 ~ 980ti (according to some 3dmark 'leaks' it was 10% slower than 980ti).
(http://www.3dmark.com/3dm11/11061015)

So take rumour 1 and rumour 3 ===> polaris 10 ~ 1070.

So unless 1080 is 40% faster than 1070, the 40% bigger die size and more expensive memory gives estimated 20% on top of polaris 10.

To put that in perspective:
That would be Hawai ~ a 980ti

We dont' know the real data until we have them, but I honestly don't see pascal doing that well if and only if AMD rumours and NVIDIA rumours are in sync with each other.

 

[linkgoron]

On an unrelated note, can nvidia dictate the terms of initial benchmarking?

What I'm asking is - can nvidia dictate (i.e. by contract of whatever) that sites can only use reference 980tis instead of superclocked/lightning/whatever variants when initially reviewing Pascal?

 

[Glo.]

wow so optimistic about AMD and pessimistic about Nvidia

what about something more unbiased ?

1/ I highly doubt that GP104 will have higher TDP than GM204. Nvidia won't make this mistake. It should will be 150~180W for the top SKU with GDDR5X

2/ I highly doubt that Polaris 10 will have 3072 GCN cores. All leaks go for 2560, much more reasonable if 232mm2 is right

If we take into account 1+2, it will change completely the equation and it will be much closer to the reality...

Around 200W does not mean "more than". I don't know how people can understand it that way. 180-200W. That is absolutely realistic estimate based on what we know about TSMC process. People forget that this is 1.48 GHz boost clock. Most of 300W TDP of GP100 comes from that high boost clock.

There was only one "leak" about GPU with 36 CU's(2304 cores) clocked at 800 MHz. It does not mean you are wrong, just would be hard if the rumored 4 SKUs of P10 would be cut down from only 2560 GCN core GPU. 3072 - sure. 4 versions of GPUs with both GDDR5X and GDDR5? A little bit not logical. But not impossible.

 

[Adored]

On an unrelated note, can nvidia dictate the terms of initial benchmarking?

What I'm asking is - can nvidia dictate (i.e. by contract of whatever) that sites can only use reference 980tis instead of superclocked/lightning/whatever variants when initially reviewing Pascal?

Of course, Nvidia and AMD can make any demand they want. Whether or not the site agrees to it depends on how much they value future pre-release cards, information and ad revenue.

It's more like a battle of wills though, some sites are just pure shill and cave-in easily, while others like AT can, due to their size, tell both companies where to shove their reviewer guides.

 

[Kris194]

On an unrelated note, can nvidia dictate the terms of initial benchmarking?

What I'm asking is - can nvidia dictate (i.e. by contract of whatever) that sites can only use reference 980tis instead of superclocked/lightning/whatever variants when initially reviewing Pascal?

They always use stock cards when they compare them to the other cards. It would be far from being fair to compare overclocked GPU with stock GPU.

 

[tential]

On an unrelated note, can nvidia dictate the terms of initial benchmarking?

What I'm asking is - can nvidia dictate (i.e. by contract of whatever) that sites can only use reference 980tis instead of superclocked/lightning/whatever variants when initially reviewing Pascal?

They give a guide. Same as amd.
How closely reviewers follow these guides and how detailed they are I don't know.

My guess is that Nvidia will send a review kit to use with the pascal cards and reference maxwell cards if that's what they want. And that they will be extremely tuned to current benchmarked game lineups.

I would be extremely surprised if Polaris had a clean sweep win of benchmarks at all even if it ended up being the faster chip long term or even just a couple months after the launch suite of games.

 

[xpea]

Most of 300W TDP of GP100 comes from that high boost clock.

no. the 300W TDP comes from the FP64 ALUs that consume much more FP32. That's why Titan has higher clock in gaming mode than in compute.
The 4 x 40Gb/s NVLinks cost also some power.
Moreover, Tesla TDP is specified as an absolute maximum value.

Outside HPC FP64 scenario and without NVLinks (or maybe only one for SLI), Geforce Gaming Pascal will have much lower TDP / higher efficiency.

 

[ShintaiDK]

wow so optimistic about AMD and pessimistic about Nvidia

what about something more unbiased ?

1/ I highly doubt that GP104 will have higher TDP than GM204. Nvidia won't make this mistake. It should will be 150~180W for the top SKU with GDDR5X

2/ I highly doubt that Polaris 10 will have 3072 GCN cores. All leaks go for 2560, much more reasonable if 232mm2 is right

If we take into account 1+2, it will change completely the equation and it will be much closer to the reality...

Agree. The hype train is at hyperspeed for Polaris, while the same people are busy downgrading Pascal.

 

[Adored]

Agree. The hype train is at hyperspeed for Polaris, while the same people are busy downgrading Pascal.

You would be surprised if Polaris beats the 390 on average? Nice ninja edit btw.

 

[ShintaiDK]

You would be surprised if Polaris beats the 390 on average? Nice ninja edit btw.

Yes I would. 232mm2, 256bit 6Ghz GDDR5 ~135W.

And before you mention GCN improvements, look at Tonga. 380X performance with 2048SP 5.7Ghz 256bit GDDR5. not exactly something to write home about. And DX12 Async dreams? It seems Async is very bandwidth heavy, hence why it doesn't work out well for Tonga either.

Good luck with the Polaris dreams. You can make another youtube about that

 

[Adored]

If you believed Polaris 10 wouldn't beat the 390 you wouldn't have hastily removed it from your post Shintai

Let's get real here, it'll be 980/390X minimum and if that's all it'll be a real disappointment. Beating the 980 Ti would make it the greatest GPU of all time, I don't see that happening either.

 

[Grooveriding]

Both cards will be disappointing performance improvements going on what we know so far. GP104 has the same number of transistors as GM200, any performance improvement will be coming from architecture and potential clockspeed improvements, not packing more transistors into the same area. Still, Maxwell managed to deliver about 40% more performance than full GK110 on the same process with just a 10% die size increase and removing DP functionality. The new architecture alone will probably deliver enough of an increase to justify calling the card the new flagship.

We don't know if it will clock any better than GM200 does. Intel hasn't gotten any consistent clock speed improvements going from 22nm to 14nm that I've seen. Overclock speeds are about the same on average. How that equates to TSMC's move from 28nm to 16nm, I don't know. I don't expect it will clock much better than Maxwell, it could clock similarly, it could clock worse.

I'd be happy to see 30% improvement over GM200 and expect that to be a best case scenario. I expect it will land more between 15-25% and with driver neglect of Maxwell as new games release, improve another 10% compared to its predecessor over time.

 

[ShintaiDK]

If you believed Polaris 10 wouldn't beat the 390 you wouldn't have hastily removed it from your post Shintai

Let's get real here, it'll be 980/390X minimum and if that's all it'll be a real disappointment. Beating the 980 Ti would make it the greatest GPU of all time, I don't see that happening either.

Not at all, but you see why. We are talking about Polaris in the Pascal thread now. And it doesn't belong here.

Without GDDR5X, performance wont be there for any 256bit card.

 

[ShintaiDK]

Both cards will be disappointing performance improvements going on what we know so far. GP104 has the same number of transistors as GM200, any performance improvement will be coming from architecture and potential clockspeed improvements, not packing more transistors into the same area. Still, Maxwell managed to deliver about 40% more performance than full GK110 on the same process with just a 10% die size increase and removing DP functionality. The new architecture alone will probably deliver enough of an increase to justify calling the card the new flagship.

We don't know if it will clock any better than GM200 does. Intel hasn't gotten any consistent clock speed improvements going from 22nm to 14nm that I've seen. Overclock speeds are about the same on average. How that equates to TSMC's move from 28nm to 16nm, I don't know. I don't expect it will clock much better than Maxwell, it could clock similarly, it could clock worse.

I'd be happy to see 30% improvement over GM200 and expect that to be a best case scenario. I expect it will land more between 15-25% and with driver neglect of Maxwell as new games release, improve another 10% compared to its predecessor over time.

We do know it will clock better. Just look at the P100(Gp100) clocks. Up to 40% higher than GM200 counterpart.

 

[Grooveriding]

We do know it will clock better. Just look at the P100(Gp100) clocks. Up to 40% higher than GM200 counterpart.

Isn't that with 300W power consumption though ? 980ti TDP is 250. I can clock one of my cards to 1600 on the core and it uses 350W+ This is why I don't think clocks will be much better.

 

[ShintaiDK]

Isn't that with 300W power consumption though ? 980ti TDP is what, 220 or 240 ? I can clock one of my cards to 1600 on the core and it uses 350W+ This is why I don't think clocks will be much better.

Its Tesla cards, forget your Geforce as compare.

M40 is 250W, 948Mhz base, 1114Mhz Boost.

P100 is 300W, 1328Mhz base, 1480Mhz Boost.

P100 also comes with NVLink(Intels Omipath adds 15W to compare with KNL).

 

[Grooveriding]

I checked. 980ti is also 250W. So no difference there. I think we are just seeing the same silly season expectations. I've seen the same here before every mid range die launch as flagship from nvidia. People thought 680 would be huge performance and people thought the same of 980. 680 was mediocre and 980 was pathetic.

If it's space age technology and somehow delivers something amazing then great. But realistically, it's likely going to be a very marginal performance improvement. Probably a 1200 base / 1300 boost clock setup. Another power consumption marketed release with less focus on absolute performance. There is not enough die there to give something meaty.

 

[Adored]

Its Tesla cards, forget your Geforce as compare.

M40 is 250W, 948Mhz base, 1114Mhz Boost.

P100 is 300W, 1328Mhz base, 1480Mhz Boost.

P100 also comes with NVLink(Intels Omipath adds 15W to compare with KNL).

P100 also has (surprisingly lower clocked) HBM2, which none of the previous Tesla's had.

 

[linkgoron]

They always use stock cards when they compare them to the other cards. It would be far from being fair to compare overclocked GPU with stock GPU.

If stock cards are not representative of the actual cards sold, I don't see why it's not "fair". The consumer should get benchmarks which are representative of the real world performance of the card.

If the average card (that a consumer bought) has 15-20% more FPS, I think that it's completely fair - especially if the reviewed cards are also OCd as part of the review. Obviously it's in nvidia's best interest to have stock 980tis.

http://techreport.com/review/28685/geforce-gtx-980-ti-cards-compared/5

These cards much better (out of the box) than reference. Also, as others have noted in other threads, the 980ti is basically recommended over the Fury X @4k because of its great OCing ability (and rightfully so).

Of course, Nvidia and AMD can make any demand they want. Whether or not the site agrees to it depends on how much they value future pre-release cards, information and ad revenue.

Yes, that's why I specifically used the word "dictate".

Anyway, I was just wondering as it would obviously make the new cards look much better if they're benched against stock 980ti, as for some reason I have a feeling that the 1080 won't be earth-shattering (I believe that it will be something like 15%-20% better than the stock 980ti, but that's just a feeling).

 

[Arachnotronic]

I checked. 980ti is also 250W. So no difference there. I think we are just seeing the same silly season expectations. I've seen the same here before every mid range die launch as flagship from nvidia. People thought 680 would be huge performance and people thought the same of 980. 680 was mediocre and 980 was pathetic.

If it's space age technology and somehow delivers something amazing then great. But realistically, it's likely going to be a very marginal performance improvement. Probably a 1200 base / 1300 boost clock setup. Another power consumption marketed release with less focus on absolute performance. There is not enough die there to give something meaty.

You will be pleasantly surprised, IMO. FinFETs should enable a big frequency uplift over 28nm planar.