NVIDIA Pascal Thread

[Head1985]

Gp104 will be like fermi...if 1080 have only 1920SP.it will pack useless FP64 units that only increase TDP and die size.

This is big chance for AMD...

Question is why the hell Nv will doing that?

 

[rainy]

- Some stock (and OCed) Geforce GTX 980 Ti and Fury X numbers, for reference:

- Geforce GTX 1080 (?):

~30% faster than Geforce GTX 980 Ti / ~68% faster than Geforce GTX 980 (stock clocks).

That OC 980 TI running at only 1311 MHz - we have owners of non-reference 980 Ti who run them at 1.4-1.5 GHz.

Btw, 8959 to 7489 it's not even 20 percent, I wish to know how you get circa 30 percent.

 

[coercitiv]

With 1920SP Gp104 should be only 200mm2.
If true then will Gp104 have Fp64 like Tesla.This is pretty bad for Nv to have Gp104 100% gaming SKu with half FP64 units.

I guess we'll know soon enough, although that Drive PX2 did light a question mark a while back.

What Pascal chip goes into the Drive PX2?

I'm asking because it occurs to me, not only there isn't going to be a gaming focused big chip this generation, but it may just be that all the chips in the Pascal line are HPC focused this time around.

 

[JDG1980]

It seems unbelievable that GP104 would actually have fewer shaders than GM204. Where are all these extra transistors going? GP104 should have a transistor count about as high as GM200, and it uses a narrower memory bus so it saves some die space on that front. Why can't GP104 at least match GM200's shader count (3072)?

 

[Sweepr]

That OC 980 TI running at only 1311 MHz - we have owners of non-reference 980 Ti who run them at 1.4-1.5 GHz.

Btw, 8959 to 7489 it's not even 20 percent, I wish to know how you get circa 30 percent.

Total score is irrelevant, different systems. You have to look at the graphics/GPU score: 10.102 vs 7779

BTW I doubt VRWorld has any insider info, they probably guessed GP104 = half GP100 and came up with '1.920 CUDA cores'.

 

[USER8000]

It seems unbelievable that GP104 would actually have fewer shaders than GM204. Where are all these extra transistors going? GP104 should have a transistor count about as high as GM200, and it uses a narrower memory bus so it saves some die space on that front. Why can't GP104 at least match GM200's shader count (3072)?

More complex shaders??

 

[R0H1T]

Gp104 will be like fermi...if it have only 1920SP.it will pack uselles FP64 units that only increase TDP and die size.

This is big chance for AMD...

Question is why the hell Nv will doing that?

No dedicated compute is as important as anything else on DX12, it'd be a huge mistake if Nvidia were to neglect that part of the equation.

 

[96Firebird]

Btw, 8959 to 7489 it's not even 20 percent, I wish to know how you get circa 30 percent.

Stock 980 Ti's graphics score from LegitReviews: 7779
Unknown card's grahpics score from leak: 10102

10102/7779 = 1.2986, or ~30% higher.

Edit - Waay late...

 

[Head1985]

Fp64 are for DP not DX12

 

[csbin]

1070 7.5GB？

 

[Head1985]

old fake leak

 

[JDG1980]

No dedicated compute is as important as anything else on DX12, it'd be a huge mistake if Nvidia were to neglect that part of the equation.

Compute in general, yes. But virtually nothing except HPC applications needs FP64. There's a reason that AMD and Nvidia both limited strong FP64 to their flagship cards for the 28nm generation (and sometimes skipped out on it even then, like with GM200 and Fiji).

 

[renderstate]

What does "dedicated compute" even mean?

 

[Timmah!]

With 1920SP Gp104 should be only 200mm2.
If true then will Gp104 have Fp64 like Tesla.This is pretty bad for Nv to have Gp104 100% gaming SKu with half FP64 units.

This. If GP104 is 2048 CC, compared to P100: 2048/3840 = 0,53. Die size: 320/610 = 0,52. However P100 has + 1920 FP64 units, which we assume wont be there with GP104...well, utter majority of them anyway. Neither the NV link stuff, which i guess takes additional die space too.

If it turns out to be 2048 CC part, then no buy from me for sure. Even if its clocked to almost 2GHz to make it "significantly" faster than 980Ti. I could OC it myself, but cant add shaders.

 

[R0H1T]

Compute in general, yes. But virtually nothing except HPC applications needs FP64. There's a reason that AMD and Nvidia both limited strong FP64 to their flagship cards for the 28nm generation (and sometimes skipped out on it even then, like with GM200 and Fiji).

What about VR, surely there must be something out there worth dedicating FP64 units for?

What does even "dedicated compute" mean?


Sent from my iPhone using Tapatalk

Let's just say compute for (no further) argument's sake 😀

 

[Det0x]

Total score is irrelevant, different systems. You have to look at the graphics/GPU score: 10.102 vs 7779

BTW I doubt VRWorld has any insider info, they probably guessed GP104 = half GP100 and came up with '1.920 CUDA cores'.

Since pretty much all 980ti sold recently have been after-market versions, why should we not compare vs that instead ?

If we use your owned linked numbers: (980ti at a cossy 1311/2003)

graphic score:

980ti = 9201
1080 = 10102

10102/9201 = ~9% -> within range of previews leaked 980ti with higher clocks

Let me ask you something else, if you want to compare a 290x with some other card. Which version is the best to use if you want "real numbers" ?

The choking stock version, or the after market one that people actuality owns ?

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 silicon yet and a recent article on SA could point to why.

You are getting 2 years old performance numbers at the same price, a good buy ?

 

[CakeMonster]

Gp104 will be like fermi...if 1080 have only 1920SP.it will pack useless FP64 units that only increase TDP and die size.

This is big chance for AMD...

Question is why the hell Nv will doing that?

Why NOT?

They are going to need to milk 16/14nm for 4-6 years with gradual performance increases.

 

[Sweepr]

Since pretty much all 980ti sold recently have been after-market versions, why should we not compare vs that instead ?

Because there will be aftermarket Geforce GTX 1080's as well, unless for some reason you have to buy the reference model.

Also because this card probably targets GM204 and Fermi/Kepler users, if you own an OCed Geforce GTX 980 Ti you might want to wait for GP100/GP102 or GP104 price cuts.

 

[Glo.]

So it will be 1/2 of the 300W GP100, with GDDR5X which will use more power than HBM2, and with extremely high core clocks. I think 185W-200W is the power target where we are looking.

 

[Head1985]

Why NOT?

They are going to need to milk 16/14nm for 4-6 years with gradual performance increases.

Gp104 with size 200mm2 and without FP64 will have same performance as 300+mm2 GP104 with FP64units.

1-Nv is crazy and GP104 only have 1920-2048SP like GTX980 and with 300mm2 SKU.300mm2 will have worse yields than 200mm2.They must be just Crazy to do this.
2-3d mark read that boost clock bad and GP104 will have 2560+ SP with no FP64 units like maxwell and clocked at 1400-1500mhz
3-its all fake

 

[nvgpu]

http://www.geforce.com/whats-new/articles/nvidia-livestream-may-6th-at-6pm-pst

 

[xpea]

http://www.geforce.com/whats-new/articles/nvidia-livestream-may-6th-at-6pm-pst

They go public tomorrow. interesting

 

[RussianSensation]

It seems unbelievable that GP104 would actually have fewer shaders than GM204. Where are all these extra transistors going? GP104 should have a transistor count about as high as GM200, and it uses a narrower memory bus so it saves some die space on that front. Why can't GP104 at least match GM200's shader count (3072)?

That's why when P100 came out and some people started saying GP104 will have 3000-4000 CC, and GP102 would have 5000-6000 CC, with GP104 rumored to have 1920-2560, it clearly shows those predictions were more realistic for Volta. A 3840 CC 1080Ti with 1.6-1.7Ghz clocks is more than enough for a generational leap over 980Ti.

Anyway, I predict NV PR will ensure that 1070/1080 are compared against reference 980Ti cards, essentially neutering 980Ti that people actually purchase by 15-20%. We'll need to know 1080's overclocking % to see if it's actually truly 25-30% faster than a 980Ti since 980Ti can overclock from ~ 1202 to 1500+.

I wish to understand logic of some Nvidia fans: GTX 980 Ti was released less than year ago (2 June 2015) at 649 dollars price.
If 1080 is 25 percent faster and will cost 600-650 dollars, it's really worth to get it?
Nvidia for sure will release significant faster card in the next 9-12 months - personally, I would prefer 50-60 percent more performance having such strong card like 980 Ti.

Per NV's Investor Day 2016 slides, 70% of NV's install base is on pre-Maxwell architectures. Per AMD, less than 10 million have 970 or above. That leaves a big market for gamers who will want to upgrade to GP104. If you aren't in a rush and have a decent GPU, then of course it's better to wait for Big Pascal ($550 980 got owned by $650 980Ti 8 months later). Someone with an old GTX470/480/580/670/680/690/770/780/780Ti though probably won't wait another 9-12 months. 980Ti users can sell their card for $475-500 this month and roll-over the resale value towards say a 20-25% faster 1080. That's why NV launched 1080 for $550+ first. It's a brilliant strategy that seems to have 0 weaknesses. The only way NV will fail with this strategy is if AMD launches big die first and mid-range 2nd; but AMD isn't even launching mid-range first. Polaris 10 is mainstream. That gives NV 100% control of their 1st mover advantage strategy and dictate pricing.

 

[Glo.]

Also it might mean this: GTX 1070 is cut down to 1792 CUDA cores, and uses GDDR5 memory, OR it is 1920 CUDA core GPU with GDDR5 memory.

 

[tential]

What are you suggesting will be the best option for midrange owners that have 970s/290s?