Nvidia BigK GK110 Kepler Speculation Thread

LittleJ said:
Well Chuck seems to think GK110 will indeed seem to find its way to GeForce products, likely in some salvaged die low yielding form....
Click to expand...

I see Chuckie is still posting his tired old FUD. Somethings never change with him.

Since his "Fermi was going to be a bad gaming GPU because it was had compute features" articles turned out to be completely wrong there is no reason to believe that his latest attack piece has any validity.

Oh and his 600 mm2 size is also wrong by 80 mm2. It is more like 520 mm2.

http://forum.beyond3d.com/showpost.php?p=1643040&postcount=111
 
The majority of die size estimates based on die shots are from slightly to a lot off base. Like everyone wanted to "know" that GK104 is 340-350-360mm2 or whatever else based on a die shot.

GK110 die area should be equal or slightly higher than 550mm2. He's not stating any exact die size either; it's somewhat bigger than GF110 and smaller than GT200@65nm.
 
LittleJ said:
Well Chuck seems to think GK110 will indeed seem to find its way to GeForce products, likely in some salvaged die low yielding form....

http://semiaccurate.com/2012/10/15/will-nvidia-make-a-consumer-gk110-card/

If so then GK114 will likely be a just a yield optimized, perhaps higher clocked GTX 760 to compete with HD8870 Oland GPU.
Click to expand...

Says the article :

"Will Nvidia make a consumer GK110 card? There has been a lot of speculation about Nvidia’s GK110 GPU, specifically whether it will end up on the desktop or not. The short answer is that yes, it will."

Then the rest is a load of tabloid crap. I'm sure it will end up on desktop but it's possible no geforce is sold, there could be one or two quadro models, perhaps one model disabled down to 256bit and clocked low for low TDP as they did with GF100.

There can still be a geforce : maybe the GK114 is big and is GTX 780, then GK110 is used as a salvage part and is GTX 765 which fits well with the tone of the article.
 
Blazkowicz said:
Says the article :

"Will Nvidia make a consumer GK110 card? There has been a lot of speculation about Nvidia’s GK110 GPU, specifically whether it will end up on the desktop or not. The short answer is that yes, it will."

Then the rest is a load of tabloid crap. I'm sure it will end up on desktop but it's possible no geforce is sold, there could be one or two quadro models, perhaps one model disabled down to 256bit and clocked low for low TDP as they did with GF100.
Click to expand...

He states desktop. Desktop stands for GeForce and Quadro for workstation.

There can still be a geforce : maybe the GK114 is big and is GTX 780, then GK110 is used as a salvage part and is GTX 765 which fits well with the tone of the article.
Click to expand...

Can I use that sentence as a signature to have material to laugh at every time I see it? I mean please it would be the most idiotic move I would see from any IHV for decades now.

If GK110 would be not worth it they'll simply not release it all for desktop. To suggest that they'll release a 7.1b/550mm2 die severely castrated below a performance SKU, where it would be times cheaper and easier to use a GK114 salvage part is plain and simple impossible.

I'll fetch in the meantime the popcorn for the doom and gloom speculations.
 
From a post by sontin on AnandTech forums:

sontin said:
BTW: A german workstation seller has listed specs of K20 which could or could not be true:
13 SMX | 700MHz | 320bit.
Click to expand...
(I can't find the seller at the moment. EDIT: http://www.cadnetwork.de/konfigurator/gpu_rackserver_proviz_g13/system=88, thanks Ailuros!)

If true, that's 13% more theoretical SP FP performance and (if 6 Gbps memory is used) 25% more bandwidth than the 680. If a GeForce version is even further cut down, it would seem that there's almost no point in releasing it unless they are planning to pull what Blazkowicz mentioned and put it at or below the 780.

On the other hand, the Tesla/Quadro cards this generation are clocked a lot lower than the corresponding GeForce cards with the same CCs, so a GK110 GeForce may be clocked considerably higher than 700 MHz, even if it is cut down more in terms of units.

On another note, this seems to contradict rumors of Tesla K20 memory size of a multiple of 6 GB, unless there will be multiple K20 variants (in terms of SMXes, bus width, etc. not just memory). EDIT (the same edit): The link shows this K20 is 5 GB.
 
Last edited by a moderator:
The K10s are clocked at 745MHz or else ~19% lower than each GK104 core on a GTX690. The latter is rated by NV with a 300W TDP, but I can't find anywhere a TDP for the K10.
 
Blazkowicz said:
I think the K10 is 225W, I find echo of it on the web including this article dated from June.
http://www.pcper.com/category/tags/k10
The teraflops number matches with the link above.
Click to expand...
Yep, I don't know why many places list it as 250 W.

NVIDIA's board specification PDF of the K10 lists a 225 W board power.

From Heise Online: "Final specifications of Nvidia's Tesla GPU unveiled K20 with GK110" (original).

2496 CCs @ 705 MHz, 5 GB 5 Gbps GDDR5 on a 320-bit bus, 225 W.
 
Last edited by a moderator:
Still, this K20 based "tesla " workstation card look low on SP ( DP are xtremely good with 1.17Tflops at 1/3 rate so ) .. it is anyway really far of the K20 specs showed with 4.5/1.5Tflops .. ( need something like 900mhz core clock for achieve this ). But at allready 225W for this one at 705mhz, im a bit affraid of the TDP of a geforce based GK110 at 900+mhz .

So i dont know if this is the high model workstation or instead of cut Cuda cores they have cut much the core clock on this model shown.

( strange too, to see this card in configuration on a shop site, when no annonce have been made by nvidia about a release ( who should come in 2 months ( i had a half december in mind, but could be wrong )
(first time i see it before Nvidia release his Tesla sku officialy )

At this SP/DP rate this is not far of the actual competition (AMD 7970's based FirePro ) or the future Intel one's ( who annonce perf at more of 1Tflops DP )...

Outside that, for the 200gb/s, it look normal taking in account the memory is ECC ( so performance on real bandwith are lowered )
 
Last edited by a moderator:
iMacmatician said:
From a post by sontin on AnandTech forums:

(I can't find the seller at the moment. EDIT: http://www.cadnetwork.de/konfigurator/gpu_rackserver_proviz_g13/system=88, thanks Ailuros!)

If true, that's 13% more theoretical SP FP performance and (if 6 Gbps memory is used) 25% more bandwidth than the 680. If a GeForce version is even further cut down, it would seem that there's almost no point in releasing it unless they are planning to pull what Blazkowicz mentioned and put it at or below the 780.

On the other hand, the Tesla/Quadro cards this generation are clocked a lot lower than the corresponding GeForce cards with the same CCs, so a GK110 GeForce may be clocked considerably higher than 700 MHz, even if it is cut down more in terms of units.

On another note, this seems to contradict rumors of Tesla K20 memory size of a multiple of 6 GB, unless there will be multiple K20 variants (in terms of SMXes, bus width, etc. not just memory). EDIT (the same edit): The link shows this K20 is 5 GB.
Click to expand...

First gen Fermi Tesla cards had gtx470 spec a, i.e. lower than the gtx480. If GK110 isn't coming as a geforce card until March, they have plenty of time to do another respin and get higher yields with more functional units.
 
lanek said:
Outside that, for the 200gb/s, it look normal taking in account the memory is ECC ( so performance on real bandwith are lowered )
Click to expand...
No that figure is without taking ECC into account so it is indeed quite low (25% lower than FirePro W9000 and way lower than Xeon Phi will be) since it is apparently only using 5 of the 6 memory channels. Effective memory bandwidth (with ECC) is lower than that of a GTX 680 (without ECC).
 
mczak said:
No that figure is without taking ECC into account so it is indeed quite low (25% lower than FirePro W9000 and way lower than Xeon Phi will be) since it is apparently only using 5 of the 6 memory channels. Effective memory bandwidth (with ECC) is lower than that of a GTX 680 (without ECC).
Click to expand...

I was outline the words on the article about " 200GB's lie, the bandwith is lower than that in reality " .. saying this is normal, ECC lower performance, and the 200gb/s will never been reached with ECC .

Outside than that, this card dont look to be the "High end Tesla " ... 320bits memory controller ( should be 384bits on full model ), 5Gb ( should be 6gb ), and 15SMX annonced ( 13 SMX here ).
( but at the same time, 200mhz more ( ~ 900mhz ) look to be enough for reach the Tflops/SP annonced for the K20 ) . ( maybe my calculation is bad )

Now.. maybe this explain why we see this card enter the market by some backdoor.. ( Cad material configurator ).. I dont say the full model will be like that, but i still dont remember seeing Nvidia let this type of things happend with the " big models ".
 
Last edited by a moderator:
You must log in or register to reply here.

Similar threads

A
Nvidia: BlueField family of DPUs for data centers
Replies
6
Views
4K
A1xLLcqAgt0qc2RyMz0y
A
Shortbread
Next-Generation NVMe SSD and I/O Technology [PC, PS5, XBSX|S]
105 106 107
Replies
2K
Views
217K
Karamazov
K
D
Nvidia Turing Architecture [2018]
15 16 17
Replies
338
Views
99K
Kaotik
Kaotik
A
NVIDIA Files Complaints Against Samsung and Qualcomm for Infringing Its GPU Patents
7 8 9
Replies
176
Views
71K
xz321zx
X
L
Dazed and Computed / did the GPUs took the wrong road?
2
Replies
27
Views
10K
liolio
L
Back
Top