NVIDIA Kepler speculation thread

[Arty]

>_<

NVIDIA GeForce Kepler GK107 (Engineering Sample) 3DMark 06 Benchmark!
http://www.youtube.com/watch?v=NekA62118Yg

NVIDIA GeForce Kepler GK107 (Engineering Sample) 3DMark 11 Benchmark!
http://www.youtube.com/watch?v=2GLcG5RPX-c

Nice find. Going by those scores it has Cape Verde beat by a significant margin. That is a very good score for a supposedly tiny card. Wasnt this supposed to go up against Turks and not Cape Verde?

If it works out like that then GK106 might be competitive with Pitcairn while GK104 being competitive with Tahiti (Pro). Even if GK100 shows up late as per the latest round of rumours, Nvidia could still very well be in a good position.

*EDIT* Added the scores here in case the YT videos are taken down.

3DMARK 2005 (1024x768)
31734 3DMarks

3DMARK 2006 (1280x720)
23893 3DMarks
SM2.0 Score 9651
HDR/SM3.0 Score 11408
CPU Score 6274

3DMARK 2011
P3818
Graphics Score 3521
Physics Score 6725
Combined Score 3767
GT1 17.27
GT2 16.70
GT3 21.08
GT4 10.41
PT 21.35
CT 17.52

3DMARK VANTAGE
P16079
GPU Score 15435
CPU Score 18739
GPU Test1 47.83
GPU Test2 42.50
CPU Test1 2542.01
CPU Test2 24.39
Feature Test1 37.90
Feature Test2 5.00
Feature Test3 28.70
Feature Test4 68.89
Feature Test5 64.51
Feature Test6 69.61

RESIDENT EVIL 5
Average 111.5fps @ 1920x1080

 

[DarthShader]

http://www.youtube.com/watch?v=2GLcG5RPX-c

"big thnx to nvidia for the Engineering Sample."

Sounds legit!

 

[mczak]

Nice find. Going by those scores it has Cape Verde beat by a significant margin. That is a very good score for a supposedly tiny card. Wasnt this supposed to go up against Turks and not Cape Verde?

I would think this is the fastest version (in contrast to the vrzone one, not sure if the pcb is going to be the same - I tend to think).
The lead over HD7770 isn't that big but at least in 3dmark11 definitely there (5-10% or so).
Not sure why you think it should go up against Turks? This most likely has similar "physical stats" to CV.
I'm wondering how high clocks and power draw were? The score seems almost too good for a 384SP part with no hotclock, unless there were some other significant architecture improvements it would need like ~1100Mhz for that score (extrapolating from GF116).

If it works out like that then GK106 might be competitive with Pitcairn while GK104 being competitive with Tahiti (Pro). Even if GK100 shows up late as per the latest round of rumours, Nvidia could still very well be in a good position.

Assuming the video is real, it would indeed be good news for nvidia. I think it's fake, though.
FWIW the same guy has also videos of 3dmark05, vantage, and RE5 on gk107.

 

[Tchock]

I doubt that an engineering sample would identify itself as that instead of the more cryptic positioning names (D14M/P-numberesque etc).

so... 80% fakeness spotted!

 

[Arty]

Not sure why you think it should go up against Turks? This most likely has similar "physical stats" to CV.

It's the successor to GF108 and expected to go up against IGPs and APUs.

 

[Blazkowicz]

so GK107 is the lower 128bit part, GK106 the higher 128bit part, then GK104 and GK110 are 256bit and 384bit? (or 512bit)

assuming they don't go for useless 192bit again.

guy has an e-peen complex with i3960X and republic of gamer desktop picture on his youtube video.
esp. weird is card identifying as "(engineering sample)"

 

[ninelven]

assuming they don't go for useless 192bit again.

[strike]Why, exactly, is a 192-bit memory interface useless?[/strike] There is nothing wrong with a 192-bit memory interface.

 

[Blazkowicz]

well I wanted my cut down, cheap and low power 192bit ddr3 gf106 or gf116. (something to play with my low res). we didn't get it .
they launched a terrible, useless GT 440 instead.

the full 192bit card was also overpriced (gtx 550 ti) and crippled somewhat because giving it 1.5GB would make the better 1GB cards look bad.

I guess it was a case of, the chip is so big we have ample room for a a 192bit bus, let's put it and find some marginal use for it.

 

[Man from Atlantis]

3DMARK VANTAGE
Feature Test1 37.90
Feature Test2 5.00
Feature Test3 28.70
Feature Test4 68.89
Feature Test5 64.51
Feature Test6 69.61

not that i believe these scores
texture filtering 37.90 GT/s
pixel filtering 5.00 GP/s

so 16ROPs 64TMUs @600MHz or 8ROPs 32TMUs @1200MHz i call it fake, most likely they renamed the card via bios and the card they used probably was a GF104/114 based.. as Tchock said it must have a mystical name at this stage, this is not a intel cpu

Spoiler

EDIT: GTX460(336CCs,14ROPs,56TMUs)@775MHz
Feature Test1 37.86
Feature Test2 5.01
Feature Test3 28.94
Feature Test4 65.60
Feature Test5 76.45
Feature Test6 68.45

Spoiler

except test 5 there is no significant difference.. and its particle simulation according to kepler rumors it should be blazingly fast for physics lol..

 

[AnarchX]

I doubt that an engineering sample would identify itself as that instead of the more cryptic positioning names (D14M/P-numberesque etc).

No, there are drivers out, where GK107 is named "GK107". This would be the name, which the apps report.

But it really looks a bit to fast for a GK107.

edit:
But in 295.73, which is used in this videos (DXdiag), there is no native GK107 support.

 

[boxleitnerb]

CPU scores are way too for a 3960X. Maybe it's fake after all.

 

[CarstenS]

The non-hotclocked/1,536-ALU-thingie does not sound very impressive when related to GF104/b and the new process tech at all - at least not when you take raw FLOPS as a measurement.

 

[Psycho]

EDIT: GTX460(336CCs,14ROPs,56TMUs)@775MHz
Feature Test1 37.86
Feature Test2 5.01
Feature Test3 28.94
Feature Test4 65.60
Feature Test5 76.45
Feature Test6 68.45

Nice, that pretty much seals it (as fake).

 

[mczak]

It's the successor to GF108 and expected to go up against IGPs and APUs.

gf108 was the chip to go up against Turks - couldn't quite do it but not that much behind. CV is successor (sort of) Turks hence gk107 would be quite a natural competitor. There is no successor to gf119 (which could have Turks-like performance) as far as I can tell.

[strike]Why, exactly, is a 192-bit memory interface useless?[/strike] There is nothing wrong with a 192-bit memory interface.

I don't know if there was nothing wrong (due to the "uneven" distribution of memory in case of the 1GB equipped cards) but in any case gf106/gf116 simply do not need that much memory bandwidth (when equipped with gddr5). gf104 came in 192bit flavors, and performance was only slightly lower. Since gf106/gf116 is essentially half a gf104/gf114, a 192bit bus there is like gf104/gf114 having a 384bit bus, it's just overkill. Even more ludicrous if you look at color rops: The ROPs of gf106/gf116 could output 24 pixels/clock, whereas the rest of the chip is limited to 8 pixels/clock...

 

[Kaotik]

gf108 was the chip to go up against Turks - couldn't quite do it but not that much behind. CV is successor (sort of) Turks hence gk107 would be quite a natural competitor. There is no successor to gf119 (which could have Turks-like performance) as far as I can tell.

Not quite, CV is successor of Juniper and nothing else, while Pitcarin is successor of Barts and Tahiti of Cayman.
Anything below Juniper, including Turks, is succeeded by APUs only.

 

[mczak]

Not quite, CV is successor of Juniper and nothing else, while Pitcarin is successor of Barts and Tahiti of Cayman.
Anything below Juniper, including Turks, is succeeded by APUs only.

I don't think that's quite true. CV definitely looks like it should also be a replacement of at least the full-blown gddr5 turks solutions. Power draw, die size etc. suggest it's not meant as only a Juniper replacement.

 

[ninelven]

I don't know if there was nothing wrong (due to the "uneven" distribution of memory in case of the 1GB equipped cards) but in any case gf106/gf116 simply do not need that much memory bandwidth (when equipped with gddr5). gf104 came in 192bit flavors, and performance was only slightly lower. Since gf106/gf116 is essentially half a gf104/gf114, a 192bit bus there is like gf104/gf114 having a 384bit bus, it's just overkill. Even more ludicrous if you look at color rops: The ROPs of gf106/gf116 could output 24 pixels/clock, whereas the rest of the chip is limited to 8 pixels/clock...

What you are describing is perceived problems with implementation, not anything inherently flawed about a 192-bit memory interface.

The non-hotclocked/1,536-ALU-thingie does not sound very impressive when related to GF104/b and the new process tech at all - at least not when you take raw FLOPS as a measurement.

Oh, I don't know. I suppose that would depend on how efficient they are. Considering what Nvidia is able to accomplish with ~1.7 TFLOPS in in the GTX 580, ~3 TFLOPS could be pretty impressive indeed.

 

[Ailuros]

not that i believe these scores
texture filtering 37.90 GT/s
pixel filtering 5.00 GP/s

so 16ROPs 64TMUs @600MHz or 8ROPs 32TMUs @1200MHz i call it fake, most likely they renamed the card via bios and the card they used probably was a GF104/114 based.. as Tchock said it must have a mystical name at this stage, this is not a intel cpu

Spoiler

EDIT: GTX460(336CCs,14ROPs,56TMUs)@775MHz
Feature Test1 37.86
Feature Test2 5.01
Feature Test3 28.94
Feature Test4 65.60
Feature Test5 76.45
Feature Test6 68.45

Spoiler

except test 5 there is no significant difference.. and its particle simulation according to kepler rumors it should be blazingly fast for physics lol..

Even if they would try to sell it as a GK106 I wouldn't fall for it for other reasons

 

[iMacmatician]

From EXPreview (translated):

Biography "Kepler" graphics to be delayed release, but still in the March debut
…
Allegedly, "Kepler" card release time from the original mid-March, was postponed to the end of March, but it should not be dragged into April, the specific date, is still unable to determine, the extension of reason and shall not known. We hypothesize that this should be the poor and the TSMC 28nm process yields, its direct impact NVIDIA had to postpone the release of "Kepler".
…

 

[mczak]

What you are describing is perceived problems with implementation, not anything inherently flawed about a 192-bit memory interface.

That's right. This referred to gf106/gf116 however specifically, where it arguably was not terribly useful (except you could get somewhat decent ddr3 cards). I see no reason why gk107, a chip which supposedly should end up with a somewhat similar performance level, could make more use of a 192bit interface.

Oh, I don't know. I suppose that would depend on how efficient they are. Considering what Nvidia is able to accomplish with ~1.7 TFLOPS in in the GTX 580, ~3 TFLOPS could be pretty impressive indeed.

I suspect with such a large alu increase it will shift performance bottlenecks to elsewhere - just like AMD cards weren't really terribly limited by simds for a long time now. That's not to say a ~3TFlops gk104 couldn't be quite impressive, but you certainly shouldn't expect an increase corresponding to the peak flop increase neither.
I don't quite agree though with CarstenS that it wouldn't be impressive for a successor of gf104. I think more than doubling alu capacity is indeed quite impressive (if they can achieve the clocks do actually exceed the factor 2).