NVIDIA Pascal Thread

[Ajay]

I thought that someone compared GM200 and GP100 per shader performance and found that it was 35% higher in Pascal. Of course, I just accidentally erased all my brower's history (new version on CCleaner - d'oh!). So I have no way to back this. If it's true, it more than makes up for the lesser amount of shaders. If it's not - I just have egg on my face.

 

[Silverforce11]

GTX 980 Ti has 2816 shaders and can clock to above 1400 MHz on the best aftermarket cards (>1500 MHz for those that include closed-loop water coolers). If GP104 really has only 2560 shaders, then either shader IPC must have gone up a lot since Maxwell, or the clock gains are really stupendous, or else the performance won't be as good as many people are hoping.

All the usual shill sites that threw OC 970, 980 and 980Ti against AMD's stack, on Pascal debut, it'll be compared to reference Maxwell.

Though, IPC will go up, better wavefront layout, modern games are optimized for 64 wavefront that Maxwell suffers at slightly (Kepler tanks hard). Likewise, compute heavy games will see a nice gain with Pascal with their GCN-like fine-grained preemption capable uarch.

It'll come down to top clockspeeds, because ultimately, at least we can see through the fud and compare OC vs OC.

 

[Silverforce11]

I thought that someone compared GM200 and GP100 per shader performance and found that it was 35% higher in Pascal. Of course, I just accidentally erased all my brower's history (new version on CCleaner - d'oh!). So I have no way to back this. If it's true, it more than makes up for the lesser amount of shaders. If it's not - I just have egg on my face.

No, that was Maxwell's gain over Kepler.

Since Pascal is Maxwell+, I don't think it will be that high. Maybe 20% in modern DX11 games or DX12 games.

 

[tajoh111]

And people say that expectations about Polaris GPUs are high...

It actually not that unrealistic, basically with that estimate, your adding about 54% more transistors and getting 45% more performance which is pretty realistic. Remember he said gtx 980, not gtx 980 ti.

From the latest rumors and going by the 380x performance, for AMD to match the gtx 980 ti, they need a 90% increase in performance for what is about 26-38% more transistors. That's pretty darn difficult.

If AMD get the same increase in performance as the above Nvidia estimate, basically it's performing around an r9 390 level or 290x, actually a tad less down to r9 290 level. Polaris 10 is only 232mm2 and should only be between 6.3-6.9 billion transistors or a 26-38% increase in transitors, hence the 21-31% percent increase over Tonga. Of course clocks can add another 10-20 percent but the same can be said of nvidia as well for pascal. With a 15 addition in performance due to clocks, it performs between an r9 390 and a 390x with performance closer to the r9 390.

The above estimates were calculated from techpowerups charts

 

[Silverforce11]

@tajoh111
14nm FF is 2.2x more denser, ie, more transistors per area than 28nm.

232mm2 on 14nm FF is equivalent to ~510mm2 on 28nm. Quite a bit bigger than Hawaii (390X).

 

[renderstate]

@tajoh111
14nm FF is 2.2x more denser, ie, more transistors per area than 28nm.

232mm2 on 14nm FF is equivalent to ~510mm2 on 28nm. Quite a bit bigger than Hawaii (390X).

I/O interfaces really don't scale that well with new processes, in practice you are not going to get a 2x scaling factor over the entire GPU.

 

[Silverforce11]

I/O interfaces really don't scale that well with new processes, in practice you are not going to get a 2x scaling factor over the entire GPU.

True. The 2.2x transistor density may end up being ~2x equivalent to 28nm when factoring in I/O. That still puts it bigger than the Hawaii equivalence.

Add uarch and clock speeds, it should be faster than Hawaii... UNLESS AMD really went all out to chase perf/w with lower clocks.

I'm not excited about getting 390X performance at ~100W, but I can see the potential sell for the mainstream segment for that. Lots of el-cheapo OEM systems have terrible PSUs and worse airflow.

In fact, thinking about it, GP104 and Polaris 11/10 are both very MEH for someone who already owns a good GPU...

 

[Ajay]

No, that was Maxwell's gain over Kepler.

Since Pascal is Maxwell+, I don't think it will be that high. Maybe 20% in modern DX11 games or DX12 games.

Yeah, it was an estimated 30% over at chiphell - so, probably worth nothing.
Maxwell and Pascal are the general architectures - so NV could have improved performance per CC quite a bit if they chose to (they've had plenty of time to do so) - we'll just have to wait and see.

 

[xpea]

I'm surprised also, since they recently brought out the 75W 950 models. I thought that was an attempt to be more competitive with Polaris. Do you have a new SKU for only 1 month?

They will simply charge more for 1060Ti

970 was already terrible cutdown.I cant imagine how Nv can screw GTX1070 even more.

I think Nvidia did incredibly well with GTX970 as it's the most successful GPU of this generation (look a steam statistics)

 

[USER8000]

True. The 2.2x transistor density may end up being ~2x equivalent to 28nm when factoring in I/O. That still puts it bigger than the Hawaii equivalence.

Add uarch and clock speeds, it should be faster than Hawaii... UNLESS AMD really went all out to chase perf/w with lower clocks.

I'm not excited about getting 390X performance at ~100W, but I can see the potential sell for the mainstream segment for that. Lots of el-cheapo OEM systems have terrible PSUs and worse airflow.

In fact, thinking about it, GP104 and Polaris 11/10 are both very MEH for someone who already owns a good GPU...

People are also comparing Polaris which is probably not designed for significant DP performance with Hawaii,which is,so the latter has transistors dedicated towards things which are not relevant for gaming. The GM200 also was not designed significant DP performance like the GP104 probably is,as it was a scaled up GM204. Hawaii is more comparable in some ways to the GK110 and GK210.

Tonga apparently had a 384 bit memory controller which was never used,and supposedly had compatibility with HBM,so even Tonga probably has parts of the chip which are inactive.

In the end the 14NM LPP process is going to be the more limiting factor.

 

[Timmah!]

I thought that someone compared GM200 and GP100 per shader performance and found that it was 35% higher in Pascal. Of course, I just accidentally erased all my brower's history (new version on CCleaner - d'oh!). So I have no way to back this. If it's true, it more than makes up for the lesser amount of shaders. If it's not - I just have egg on my face.

I wonder how he did it. I tried on my own by multiplying (if thats the right word) the number of CC with the boost frequency of both Titan X and P100 and then compared the resulting number to claimed FP32 peak performance number - apparently though, this was not right the way to do. So i am truly interested how to calculate it properly.

Anyway, that way i described it, Tesla P100 seemed to have about 10 percent more performance... i mean, the peak FP32 perf number should have been 10 percent lower to proportionally fit to the CC number x boost clock comparison between P100 and Titan X.

I am shit at math though.

 

[seitur]

In fact, thinking about it, GP104 and Polaris 11/10 are both very MEH for someone who already owns a good GPU...

Kinda exected for some months already.

Real High-end GPU replacements are coming in 2017 with Vega and Nvidia equivalent.

 

[Timmah!]

Kinda exected for some months already.

Real High-end GPU replacements are coming in 2017 with Vega and Nvidia equivalent.

Yeah, but that does not really fit the narrative of Nvidia wanting the people to upgrade twice per generation.

I still have my hopes up for 1080 to be 3200 CC part, even if the current rumors say otherwise. Not that it did not happen before, but honestly if Nvidia expects/wants even the Titan X/980Ti owners to upgrade, they should not release a card with less CC than those, even if the resulting card is ultimately faster cause of higher IPC. Its a psychological thing - you dont claim your IPC improvement on the box, the CC number you do.

 

[airfathaaaaa]

Yeah, but that does not really fit the narrative of Nvidia wanting the people to upgrade twice per generation.

I still have my hopes up for 1080 to be 3200 CC part, even if the current rumors say otherwise. Not that it did not happen before, but honestly if Nvidia expects/wants even the Titan X/980Ti owners to upgrade, they should not release a card with less CC than those, even if the resulting card is ultimately faster cause of higher IPC. Its a psychological thing - you dont claim your IPC improvement on the box, the CC number you do.

they can always call volta a maxwell 2.0 of pascal and call it an upgrade..
i dont see how the CC count will actually make anyone upgrade...if the perf is there with less people wont say "oh it doesnt have 3200 cc? screw you nvidia i wont buy it"...this never happened because the majority of the people never will actually look at so deep in order to buy a card

 

[Timmah!]

they can always call volta a maxwell 2.0 of pascal and call it an upgrade..
i dont see how the CC count will actually make anyone upgrade...if the perf is there with less people wont say "oh it doesnt have 3200 cc? screw you nvidia i wont buy it"...this never happened because the majority of the people never will actually look at so deep in order to buy a card

Naturally its gotta be somewhat faster than previous gen, higher CC number by itself does not warrant anything, if it does not bring performance increase alongside it.... but as i said, its one of the parameters/properties of the GPU. Like VRAM capacity. Its usually stated on the box. Its one way to compare to previous generation. To someone, who is not hardware enthusiast, it now says its in some way inferior to what he already owns. And as you said, will this person look deeper inside to actually find out the card is faster, even though based on specs it should not be?

And then again someone like me. Lets say i own Titan X, hypothetically (i dont). But now i see the 1080 is say 25 percent faster, even though it has only 2560 shaders. Do you know what i think? If only it had 3200 of those (more than my Titan X), it would have been perhaps 40-50 percent faster. Now that would make me to want to buy it. And i know 3200 CC would have been manufacturable for Nvidia (if p100 is after all), the chip would just need to be somewhat larger (say 350mm instead of 320mm). Yet Nvidia decided that 25 percent perf increase should be good enough for me (cause somewhat better yields, higher margins for them). So yeah, in such case, personally i am inclined to think, screw you Nvidia.

 

[airfathaaaaa]

Naturally its gotta be somewhat faster than previous gen, higher CC number by itself does not warrant anything, if it does not bring performance increase alongside it.... but as i said, its one of the parameters/properties of the GPU. Like VRAM capacity. Its usually stated on the box. Its one way to compare to previous generation. To someone, who is not hardware enthusiast, it now says its in some way inferior to what he already owns. And as you said, will this person look deeper inside to actually find out the card is faster, even though based on specs it should not be?

And then again someone like me. Lets say i own Titan X, hypothetically (i dont). But now i see the 1080 is say 25 percent faster, even though it has only 2560 shaders. Do you know what i think? If only it had 3200 of those (more than my Titan X), it would have been perhaps 40-50 percent faster. Now that would make me to want to buy it. And i know 3200 CC would have been manufacturable for Nvidia (if p100 is after all), the chip would just need to be somewhat larger (say 350mm instead of 320mm). Yet Nvidia decided that 25 percent perf increase should be good enough for me (cause somewhat better yields, higher margins for them). So yeah, in such case, personally i am inclined to think, screw you Nvidia.

i never saw any box out there showcasing that would be pretty useless tho they usually showcase fancy upgrades and cc count isnt one of them

also yeah nvidia could create a 400mm2 too and have 8192cc does it make it reasonable? no ofc not.. and yes if history is to taken seriously when it comes to nvidia tacticts you can rest assure that they will do such a thing and call it an upgrade (titanx/980ti)

 

[Timmah!]

i never saw any box out there showcasing that would be pretty useless tho they usually showcase fancy upgrades and cc count isnt one of them

also yeah nvidia could create a 400mm2 too and have 8192cc does it make it reasonable? no ofc not.. and yes if history is to taken seriously when it comes to nvidia tacticts you can rest assure that they will do such a thing and call it an upgrade (titanx/980ti)

I am aware of Nvidia tactics. All i am saying, if i am owner of TitanX/980Ti, then 40-50 percent faster 3200 CC GPU would be more desirable for me (and hopefully other people) than 25 percent faster 2560 CC part (all hypothetical numbers), while it would not take anything from next Titan/1080Ti, since those would be significantly faster next year again anyway (at 600mm2).

 

[Silverforce11]

I am aware of Nvidia tactics. All i am saying, if i am owner of TitanX/980Ti, then 40-50 percent faster 3200 CC GPU would be more desirable for me (and hopefully other people) than 25 percent faster 2560 CC part (all hypothetical numbers), while it would not take anything from next Titan/1080Ti, since those would be significantly faster next year again anyway (at 600mm2).

A lot of that 600mm2 from GP100 is dedicated to FP64 & NVLink which has no use in consumer gaming. So don't expect too much out of GP100.

I would not be surprised if it's ~25-30% above the 1080.

But I agree with you, I don't upgrade for less than 50% perf gains. Preferably even higher.

 

[airfathaaaaa]

I am aware of Nvidia tactics. All i am saying, if i am owner of TitanX/980Ti, then 40-50 percent faster 3200 CC GPU would be more desirable for me (and hopefully other people) than 25 percent faster 2560 CC part (all hypothetical numbers), while it would not take anything from next Titan/1080Ti, since those would be significantly faster next year again anyway (at 600mm2).

40-50 from a enthusiast tier to a high tier? this will never happen not only because of the pr nightmare but because it will be impractical simply because the next enthusiast will need to be again on a 40-50% from the high that just got 40-50% from the previous enthusiast(purely to keep the psychological and pr effect in place) can you imagine with that rate we will reach the current node limit pretty PRETTY fast

 

[Timmah!]

A lot of that 600mm2 from GP100 is dedicated to FP64 & NVLink which has no use in consumer gaming. So don't expect too much out of GP100.

I would not be surprised if it's ~25-30% above the 1080.

But I agree with you, I don't upgrade for less than 50% perf gains. Preferably even higher.

I do expect it to be at least those 3840 shaders minus the FP64 and NV Link part. Granted, in that case it may not have those 600 mm2, perhaps 450 mm2 would be enough.

40-50 from a enthusiast tier to a high tier? this will never happen not only because of the pr nightmare but because it will be impractical simply because the next enthusiast will need to be again on a 40-50% from the high that just got 40-50% from the previous enthusiast(purely to keep the psychological and pr effect in place) can you imagine with that rate we will reach the current node limit pretty PRETTY fast

Obviously my point was, they could have make it enthusiast tier already now. In other words that third rumored SKU could have indeed been 1080Ti 40/50 percent above 980Ti sold for 980Ti price. Then 1080 could have been 980 replacement around the perf level of 980Ti... Next year completely new line-up build around GP102/100 once again 40 percent above current offering.

 

[imaheadcase]

I am aware of Nvidia tactics. All i am saying, if i am owner of TitanX/980Ti, then 40-50 percent faster 3200 CC GPU would be more desirable for me (and hopefully other people) than 25 percent faster 2560 CC part (all hypothetical numbers), while it would not take anything from next Titan/1080Ti, since those would be significantly faster next year again anyway (at 600mm2).

Don't forget 780 Ti owners like myself. I think those you mentioned, and 780TI owners are also waiting to upgrade to Pascal than wait a year. The time frame between us and this card is pretty big.

 

[xpea]

Don't forget 780 Ti owners like myself. I think those you mentioned, and 780TI owners are also waiting to upgrade to Pascal than wait a year. The time frame between us and this card is pretty big.

Exactly. Most of the users don't update to the next generation, but n+2. So Kepler cards are the main target for the Pascal upgrade path, not Maxwell owners

 

[airfathaaaaa]

Thing go bump(gate) in the night for Nvidia’s GP100 Pascal GPU

http://semiaccurate.com/2016/05/03/thing-go-bumpgate-in-the-night-for-nvidias-gp100-pascal-gpu/

 

[Stuka87]

Thing go bump(gate) in the night for Nvidia’s GP100 Pascal GPU

http://semiaccurate.com/2016/05/03/thing-go-bumpgate-in-the-night-for-nvidias-gp100-pascal-gpu/

So what does it say, since you have to be a subscriber....

 

[airfathaaaaa]

So what does it say, since you have to be a subscriber....

ah i thought posting it directly from google would have bypass it
guess they fixed it already
ill try to find another way

well only the forum post so far http://semiaccurate.com/forums/showthread.php?p=261203
as any indication hbm serious problems