NVIDIA Pascal Thread

[Silverforce11]

1920 CUDA core GPU with 1.86 GHz core clock, 10 GHz GDDR5X memory is 25% faster than reference GTX 980 Ti, and 4% faster than GTX 980 Ti Waterforce Gaming.

It also has around 1 TFLOPs of compute power more than reference GTX 980 Ti.

TDP between 165 and 200W. Because of extremely high core clocks it will be closer to 200W.

Meh, too many contradicting rumors at this point.

Personally, I don't think it's a 1920 CC SKU, the die size is big for a 16nm FF mid-range, and this won't have GP100 1:2 FP64 ratio so what's the huge die size for?

2560 SP makes much more sense, 2/3 of GP100 in FP32 performance, minus the FP64 CC, at close to half the die size.

So I'm calling BS on that VR World claim. We shall see tomorrow whether that site is worthwhile in the future for leaks/rumors.

 

[CakeMonster]

1920 cores seems strangely low, despite the clock speed argument. Unless of course they are giving us the mid-range or even half a tier below that. I can't remember (but I might be wrong) that such a reduction is usual with a node shrink? Certainly not such a large node shrink that we have now.

 

[Det0x]

http://vrworld.com/2016/04/25/16nm-nvidia-geforce-gtx-1080-leak-ahead-computex-taipei-2016/How come people did not seen this even if it was on previous pages, already?

Like i said in my previous post:

Those 1080 scores seem low to me.

According to kitguru they get a 19859 on 3dmark11 with their "ASUS STRIX Gaming GTX 980 Ti DirectCU 3" (unless I'm missing something?) vs ~19005 of the 1080.

The zotac 980ti from guru3d gets 8891 on firestrike extreme 1.1 vs 8959 of the 1080.

So basically almost equivalent (at stock at least) to some after-market 980tis.

Yes i full heartedly agree with you.. but earlier leaks have shown the very same low performance delta between 1080 and aftermarket 980ti. (like pretty much all sold 980ti are)

Its just that some people are so blinded by their own bias that they refuse to believe anything to the contrary.

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

Unknown card:
Graphics score: 27271
GT1: 127.63 FPS
GT2: 132.23 FPS
GT3: 170.97 FPS
GT4: 80.08 FPS

GTX 980 Ti Waterforce:
Graphics score: 26112
GT1: 121.41 FPS
GT2: 126.93 FPS
GT3: 166.33 FPS
GT4: 76.31 FPS

www.3dmark.com/3dm11/11162175

Score: P20625
Graphics Score: 27271

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

For reference:

*Probably different CPUs, the Physics Score suggests a slower chip than the Core i7-5960X @ 4.4GHz used by Guru3D.

You can read the the rest back on page 64 in this thread. ()

If the rumored price of 649+$ is true, i would deem the whole geforce pascal to be a utter failure, possible even GP100 too. :'( Remember we haven't seen any working p100 silicon yet and a recent article on SA could point to why. ()

 

[Silverforce11]

1920 cores seems strangely low, despite the clock speed argument. Unless of course they are giving us the mid-range or even half a tier below that. I can't remember (but I might be wrong) that such a reduction is usual with a node shrink? Certainly not such a large node shrink that we have now.

It should seem low, because it's illogical.

We know the ~die size of GP104 and the fully enabled A400 1080. ~330-340mm2.

We know GP100's SMX layout. The most cost effective and simplest modular cut is 4/6 SMX (each SM is 640 CC). Down to 2560 CC. Remove most of the FP64 units & NVLink. What kind of size would such a chip be?

A little over half the size.

1920 CC is 3/6 SMX cut of GP100, minus the FP64/NVlink on the chip, this chip will be much smaller than GP104.

 

[Silverforce11]

@Det0x
It's not a failure.

It's within the normality of a node jump. Next-gen mid-range > previous gen high-end. ~10-30% is the normal scenario.

The price is indeed a fail, but really, we all saw it coming. The shock is if it's priced <$599. heh

 

[Det0x]

@Det0x
It's not a failure.

It's within the normality of a node jump. Next-gen mid-range > previous gen high-end. ~10-30% is the normal scenario.

The price is indeed a fail, but really, we all saw it coming. The shock is if it's priced <$599. heh

If the rumored price of 649+$ is true

It was a premise of my statement

 

[Glo.]

Architecture is similar to Kepler and Maxwell, it is just the amount of instructions executed within single SM, and the amount of the cores in single SM that comes to play here. Pascal 64 CC's have the same performance as 128 CC from Maxwell. That is what I was telling you already few weeks ago. Thats why the core count of the GPUs can be lower, and thus power consumption on that basis can be lower, or much higher performance.

340mm2 is almost exactly half the size of GP100. But GP100 does not have display engines, and ROPs. It is PURE compute unit. Also GP104 has GDDR5X memory controller, and other bit and pieces that add up to die size. And most importantly, we do not know what FP64 level of performance we are looking at.

 

[CakeMonster]

If they are dialing down the specs so much (probably not if 1920c is wrong) then it would turn me off from buying because they are obviously holding back in order to increase margins. At the same time, if the upcoming bigger die is $600+ anyway, it is just too expensive to justify. So then I might as well get the current one just because its 1) faster than my current 980, 2) its the only card in the next 2 years inside my price range.

I can feel some headaches in my future..

 

[nvgpu]

https://forum.beyond3d.com/threads/nvidia-pascal-announcement.57763/page-13#post-1905723

GP100 is fully graphics capable. They aren't drawn on the diagrams, but it has display controllers, ROPs, etc.

Also ignore any rumors about the amount of CUDA cores for GP104, it will be revealed soon anyway the official count.

 

[PPB]

New node 310mm2 die marginally better than 600mm2 die on last node and people losing their heads?

Did anyone think perf/mm2 was going magically up besides the new litho process?

Sent from my XT1040 using Tapatalk

 

[Silverforce11]

https://forum.beyond3d.com/threads/nvidia-pascal-announcement.57763/page-13#post-1905723

Also ignore any rumors about the amount of CUDA cores for GP104, it will be revealed soon anyway the official count.

Right. I wish people stop saying GP100 has no ROPS.. -_-

Neither did Fermi or Kepler Tesla diagrams. But the chip has ROPs. It's also got TMUs. Things not useful for compute, but included because it's a flexible uarch designed for both.

VR World claiming 1920 CC is out of their minds. A failure of basic logic.

 

[airfathaaaaa]

Right. I wish people stop saying GP100 has no ROPS.. -_-

Neither did Fermi or Kepler Tesla diagrams. But the chip has ROPs. It's also got TMUs. Things not useful for compute, but included because it's a flexible uarch designed for both.

VR World claiming 1920 CC is out of their minds. A failure of basic logic.

back then they needed to have because they didnt had a separate derpament for hpc and consumers...

now building a card only for your hpc department and filling it with things it wont even use its just the same of saying gp104 will be a full compute monster and very basic graphic capabilities...

 

[RussianSensation]

If they are dialing down the specs so much (probably not if 1920c is wrong) then it would turn me off from buying because they are obviously holding back in order to increase margins. At the same time, if the upcoming bigger die is $600+ anyway, it is just too expensive to justify. So then I might as well get the current one just because its 1) faster than my current 980, 2) its the only card in the next 2 years inside my price range.

I can feel some headaches in my future..

True. If this 1080 card only has 1920-2048 CCs, but we also saw GP100 with 3840 CCs as a potential max Pascal, I would rather wait 12 months and get that for $650-700. The clock speeds can be usually made up by overclocking but nearly double the CCs is an insane difference. It would mean GP100/102 OC could easily ~ 1080 SLI on average. I still think 2560 makes more sense (TX had 1.5x over 980).

On one hand, I would be very impressed if 1920 CC 1080 beats 2816 CC 980Ti by 20-25%, but that would turn me off the card for sure since it's even worse milking of the mid-range by delivering SO much less against Big Pascal.

 

[renderstate]

CUDA has supported texture units for a long time since they can provide a lot of value even to HPC applications. For instance this 2009 (!!) article already discussed the potential benefits of using textures in a CUDA applications: http://www.drdobbs.com/parallel/cuda-supercomputing-for-the-masses-part/218100902
Thinking that NVIDIA is going to suddenly drop texture units on HPC parts doesn't make any sense.

CUDA also supports writing to graphics frame buffers, so even ROPs are necessary on parts like GP100. AFAIK the only really important part of the 3D pipeline not exposed to CUDA developers is the frontend (tessellation, rasterization, culling, etc..), but probably it occupies a very small part of the chip and it there is not much value in getting rid of it.


Sent from my iPhone using Tapatalk

 

[tviceman]

Right. I wish people stop saying GP100 has no ROPS.. -_-

Neither did Fermi or Kepler Tesla diagrams. But the chip has ROPs. It's also got TMUs. Things not useful for compute, but included because it's a flexible uarch designed for both.

You're probably right. TBH, GP100's increase in size relative to GK110 and GF100 is probably accounted for in the proportional increase in DP performance (moving up to 1/2 from 1/3). If GP100 was 1/3 DP like GK110, it's die size could have likely been very similar.

Nvidia was stuck with GK110 / 210 for quite awhile as it's Tesla flagship chip so they probably realized it's better to go all out early on to counter Intel as much as possible. So in that respect, GP102 rumors would end up false.

When we get GP104's specs verified then we'll be able to make good educated guesses of GP100, insofar as graphics are concerned. Based on GK110 and GM200, though, GP100 in full form will be in the ballpark of being 30% faster than GP104.

 

[Timmah!]

back then they needed to have because they didnt had a separate derpament for hpc and consumers...

gp100 will be very likely released not only as Tesla, but Quadro too. Which is going to need those ROPs and TMUs...

True. If this 1080 card only has 1920-2048 CCs, but we also saw GP100 with 3840 CCs as a potential max Pascal, I would rather wait 12 months and get that for $650-700. The clock speeds can be usually made up by overclocking but nearly double the CCs is an insane difference. It would mean GP100/102 OC could easily ~ 1080 SLI on average. I still think 2560 makes more sense (TX had 1.5x over 980).

On one hand, I would be very impressed if 1920 CC 1080 beats 2816 CC 980Ti by 20-25%, but that would turn me off the card for sure since it's even worse milking of the mid-range by delivering SO much less against Big Pascal.

Exactly this. Knowing the full 600 mm2 die has 3840 CC (and could probably be even more, if not for FP64 units), 1920/2048 CC would be borderline insulting, regardless of how they perform compared to GM200. Especially at expected 600 EUR+ price.

 

[renderstate]

The compute parts *NEED* texture units and ROPs because they have been exposed to developers in CUDA for many years now. The idea that NVIDIA HPC parts might do away with such HW units makes no sense (unless NVIDIA are suicidal ), even if they won't ever release a Quadro counterpart.


Sent from my iPhone using Tapatalk

 

[PrincessFrosty]

Thanks to everyone who replied to me, that was helpful information and kind of what I was expecting. For a while I was quite excited about the idea of a 16nm node but it looks like the early GPUs will just be mid range attempts rather than something really nice.

The opportunity cost of holding onto a 980 for another year is too high. If you sell a 980 for $350-375 and put that $350 into a new $550 card, the new card won't drop much in the first 6 months due to lack of competition from AMD. Alternatively, it should be possible to sell a 980 and step-up to the 1070 for almost free or very little $ and 1070 should be close to 980Ti, essentially a free 25-30% boost in performance.

For 4K, I'd consider 1070 SLI over 1080 (if 970 SLI vs. 980 generation was anything to go by), but best bet is to wait for benchmarks.

This is actually what I was considering, possibly going for 2 cheaper cards, although if the 1070 isn't going to be terribly much faster than a 980 it's also possible that I'll just do what I did prior to my GTX 980 which is go from one 580 to 2x580's in SLI because the price had dropped so much. A 2nd 980 will be cheap as no doubt a lot will flood the market as others look to upgrade.

What are the rumours for a roadmap post 1080, is there some kind of Titan equivalent on the way which will really absolutely crush things like the 980 in performance? I might just wait, money is way less of an issue these days so just dropping the cash on a Titan equivalent for 4k might be less of a headache.

4k on an overclocked 980 is actually very playable, a lot of games you just need to sacrifice some of the high settings like HBAO, and you're usually good, I play most of my games in 4k now and it's really just getting the bells and whistles in 4k that will make the difference.

Kind of disappointed really this gen I had high hopes and given an unprecedented node shrink I would have expected better, we've been limping along with tiny upgrades every gen more or less since the 580 which is why I had a (never before) generational gap between the 580's and the 980 despite having the flagship card from every generation before then.

 

[Sweepr]

Time for last minute predictions. Let's see who gets it right (tonight?).
Assuming similar layout to GP100 (could be different), here's mine:

- Geforce GTX 1080
GPU configuration: 1920-2560 CUDA cores (3 or 4 GPCs / 30-40 SMs enabled)
Core clock: >1.6 GHz
Memory: 8GB GDDR5X 10 Gbps - 256-bit (320 GB/s)
Performance: 15-25% faster than Geforce GTX 980 Ti / 50-60% faster than Geforce GTX 980
Price: >$599

- Geforce GTX 1070
GPU configuration: 1536 / 2048-2176 CUDA cores (3 or 4 GPCs / 24 or 32-34 SMs enabled)
Core clock: >1.5 GHz
Memory: 8GB GDDR5 8 Gbps - 256-bit (256 GB/s)
Performance: Similar to Geforce GTX 980 Ti at 1080p/1440p / 40% faster than Geforce GTX 970
Price: >$449

 

[antihelten]

Time for last minute predictions. Let's see who gets it right (tonight?).
Assuming similar layout to GP100 (could be different), here's mine:

- Geforce GTX 1080
GPU configuration: 1920-2560 CUDA cores (3 or 4 GPCs / 30-40 SMs enabled)
Core clock: >1.6 GHz
Memory: 8GB GDDR5X 10 Gbps (320 GB/s)
Performance: 15-25% faster than Geforce GTX 980 Ti / 50-60% faster than Geforce GTX 980
Price: >$599

- Geforce GTX 1070
GPU configuration: 1536 / 2048-2176 CUDA cores (3 or 4 GPCs / 24 or 32-34 SMs enabled)
Core clock: >1.5 GHz
Memory: 8GB GDDR5 8 Gbps (256 GB/s)
Performance: Similar to Geforce GTX 980 Ti at 1080p/1440p / 40% faster than Geforce GTX 970
Price: >$449

I'm willing to go a bit more specific than that:

- Geforce GTX 1080
GPU configuration: 2560 CUDA cores (4 GPCs / 40 SMs enabled)
Core clock: 1.6-1.7 GHz
Performance: 20-30% faster than Geforce GTX 980 Ti / 50-60% faster than Geforce GTX 980
Price: $550

- Geforce GTX 1070
GPU configuration: 2176 CUDA cores (4 GPCs / 34 SMs enabled)
Core clock: ~1.6 GHz
Performance: Similar to Geforce GTX 980 Ti at 1080p/1440p / 40% faster than Geforce GTX 970
Price: $400

Perhaps a bit too optimistic (especially with the prices), but hey, no guts no glory ;P.

(I have no idea what the memory setups will be for the two, so no guesses there)

 

[tviceman]

I'm going with $599 for the 1080, $399 for the 1070, and $299 for the 660 TI that comes after Polaris.

 

[Head1985]

1080 25-30% faster than 980TI
2560SP
160TMU
64Rops
256bit with GDDR5X
320-384GB/s


1070 5-10%faster than 980TI
2304SP
144TMU
64Rops
256bit with 8Ghz DDR5
256GB/s

1060TI 5-10% slower than 980TI
2048SP
128TMU
48Rops
192Bit with 7Ghz DDR5
168GB/s

 

[SteveGrabowski]

I wonder what kind of goodies they give away at these events since it's held in Austin and open to the public. I live between San Antonio and Austin and would consider the drive up to watch it if it's typical to get coupons for the 1000-series cards or game codes or anything like that.

 

[Kenmitch]

I wonder what kind of goodies they give away at these events since it's held in Austin and open to the public. I live between San Antonio and Austin and would consider the drive up to watch it if it's typical to get coupons for the 1000-series cards or game codes or anything like that.

Only one way to find out is to go to the event. Of course you know how that works out in the end. You go and you don't get anything worth a crap. You gamble and don't go and they give away nice stuff....Well at least that's how it would work out for me at least.

 

[parvadomus]

I expect around 980TI performance if only 64ROPs, they will clock it high enough to reach +/- the same GPixel/s rate.