NVIDIA Kepler speculation thread

[psolord]

GTX 680 overclocked:
http://itbbs.pconline.com.cn/diy/15052972.html

1110MHz core
7.2Gbps memory
~8% higher GPU score over HKEPC

The difference has now jumped to 86% compared to my stock 570! lol

Yes I'm being unfair comparing the top single gpu card of the next gen plus OC, with the second top card of the previous gen, but it's still darn impressive.


@all
Any speculation of how the GTX 670 will be shaped?

 

[AnarchX]

7 SMX/1344 SPs: http://semiaccurate.com/2012/02/09/there-are-two-gk104kepler-variants/ (-400 or -355 is just SKU bining number)

Clock-rate should not be much slower.

 

[psolord]

7 SMX/1344 SPs: http://semiaccurate.com/2012/02/09/there-are-two-gk104kepler-variants/ (-400 or -355 is just SKU bining number)

Clock-rate should not be much slower.

Ah great.

So according to the hkepc leaked slides, along with those 1344SPs we could be looking at something like 112 texture units and 28 rops?

Do you think Nvidia would proceed to reduce the bus to 192 bit and go for 1.5GB framebuffer for the 670 or would they do that for the 660?

How much would a 1344sp/112tmu/28rop/256bit 2GB 670 cost?

How about a 1152/96tmu/24rop/192bit/1.5GB 660?

 

[Psycho]

How far away is the GK106? haven't really heard anything about it, but there's a huge gap between the GK107 (which seems to be ~GF106 performance) and the GK104.
So will the 660 be 104 or 106 based? and then maybe GF114 will carry them through to the 700 series.

 

[CarstenS]

If nothing changed compared to fermi, the ROPs at least are fixed to memory channels, so they cannot scale them without reducing memory interface and thus frame buffer size (with equally dense chips) as well.

I think the outcome would depend on which pricepoint they want to hit.

 

[DSC]

GK107 has 4x the SP/CUDA cores of GF108(GT 430/440). 96 x 4 = 384

GK104 has 4x the SP/CUDA cores of GF114(GTX 560 Ti). 384 x 4 = 1536

Assuming GK106 has the same 4x more ratio than GF116(GTX 550 Ti), this would mean 192 x 4 = 768 SP/CUDA cores?

 

[Vardant]

http://chinese.vr-zone.com/12537/nvidia-gtx680-sli-in-3dmark11-03182012/

What happened to 1.4Ghz on air? All I see are measly overclocks that barely go beyond turbo.

That should be without raising the voltage, afaik.

 

[Devildoll]

GK104 has 4x the SP/CUDA cores of GF114(GTX 560 Ti). 384 x 4 = 1536

Does anyone know what differentiates a Kepler cuda core from a Fermi cuda core?

i mean , crudely speaking..
GF114 = < 2B transistors
GK104 = 3.5B? transistors

how does it manage to quadruple the number of cuda cores without even ending up with twice as many transistors?

 

[AnarchX]

Kepler CUDA cores run at much lower clock rates and are organized in a more compact way - the SMX.

Gamer SMX (192SPs, 16 TMUs): ~16mm² @28nm
Gamer SM (48SPs, 8 TMUs): ~20mm²* @ 40nm

*GF100 SMs were 16mm² and Anandtech was told GF104 SMs increased the size by 25%.

 

[sheepdogexpress]

I seriously doubt board power is determined by profiling.


For those wishing for a 7970 1ghz (1.2 ghz) edition to compete against the GTX680 I really don't see the purpose.

-----------------

Lets say AMD produces a factory overclocked 1.2 ghz 7970 to outperform the GTX680.

Wouldn't the TPD of 250 watts have to be raised?
If so to what a number?
If its 300 watts isn't that dual 8 pin power then.

Now what about pricing?
It would have to be higher than the current $549 and that means even lower number of units sold.

And since AMD owners claim they could over-clock the 7970 to that same 1.2 ghz why would they ever buy the factory over clocked 1.2 ghz version.

And what about the thermals that a 1.2 ghz card would produce. That would mean more/bigger/faster fans and lots of noise.

And after all this all Nvidia has to do is either release a GTX685 with higher clocks or release the upcoming GK110.

So all in all I do not see AMD releasing a factory over clocked 7970.

Totally agree here.

Binning isn't the solution either, atleast not at this point. Yields aren't the best for the 7970 at this point and early stock shortages of the 7970 compared to the 7950 show this. Binning even further would produce even smaller quantities and make the pricing of the 7970 even less justifiable. AMD isn't as strong a brand as Nvidia, at I have doubts about an overclocked 7970 selling for 650 or 699.

Making the 7970 not top shelf anymore also makes it 550 price tag less justifiable.

Also even if the card is stable at 1.1ghz right now at stock volts, whats to guarantee it will be stable 2 years from now. I have seen people running the Leo demo and it doesn't cooperate from any overclocks. People have to think of long term stability, which is probably why AMD conservatively clocked their card. 28nm is relatively new how it ages can alter what is stable or not.

Also releasing the 7970 as fast as it can, only makes them spend more on R and D. The faster a card is to its max clocks, the slower the refresh looks in comparison. This speeds up the need for a new architecture and I think both companies don't want this.

 

[Kaotik]

Binning isn't the solution either, atleast not at this point. Yields aren't the best for the 7970 at this point and early stock shortages of the 7970 compared to the 7950 show this. Binning even further would produce even smaller quantities and make the pricing of the 7970 even less justifiable. AMD isn't as strong a brand as Nvidia, at I have doubts about an overclocked 7970 selling for 650 or 699.

You completely ignore the fact that they had over half a month more time to build 7950 stock compared to 7970 stock

 

[A1xLLcqAgt0qc2RyMz0y]

You completely ignore the fact that they had over half a month more time to build 7950 stock compared to 7970 stock

Didn't you get the note. Charlie states that 28nm production was shut down in February.

 

[sheepdogexpress]

You completely ignore the fact that they had over half a month more time to build 7950 stock compared to 7970 stock

Not at all. They have been in production for similar amounts of time and even now, I see 7970 in way less stock than the 7950. They barely have enough 7970 to go around even now. 7950's are easy as pie to get though and it is going after a significantly larger market. There is a lot more 7950 to go around than 7970s.

 

[CarstenS]

Didn't you get the note. Charlie states that 28nm production was shut down in February.

My note said:
"In 2012, Chinese New Year's Day will be celebrated on January 23. The New Year festival will start from January 22 (Chinese New Year' Eve) and end on February 6 (Lantern Festival)."

 

[cal_guy]

Didn't you get the note. Charlie states that 28nm production was shut down in February.

'I want to inform you that our 28nm production is normal, and all our 28nm customers are fully aware of our production status

Charlie is wrong.

http://www.bit-tech.net/news/hardware/2012/03/08/28nm-process-changes-tsmc/1

 

[SimBy]

Shutdown was never double confirmed.

 

[A1xLLcqAgt0qc2RyMz0y]

Charlie is wrong.

http://www.bit-tech.net/news/hardware/2012/03/08/28nm-process-changes-tsmc/1

Maybe not.

TSMC, for its part, refused to comment on the claimed stoppages or planned changes to its 28nm production process, with a spokesperson stating that it is company policy 'not to comment on market rumors'.

I don't see any denial of claimed stoppages or planned changes just a refusal to answer the question.

The spokesperson did, however, add the following: 'I want to inform you that our 28nm production is normal'

Again wishy-washy words. What exactly is 'normal'. Maybe it is 'normal' now but wasn't 'normal' in February.

and all our 28nm customers are fully aware of our production status.

Again wishy-washy. What exactly is/was the 'production status' - back to running after shutdown in February.

 

[cal_guy]

TSMC is a publicly traded corporation, if a spokesman says that 28nm production is normal then you can assume that production hasn't been shut down.

 

[A1xLLcqAgt0qc2RyMz0y]

TSMC is a publicly traded corporation, if a spokesman says that 28nm production is normal then you can assume that production hasn't been shut down.

Publicly traded on the Taiwan Stock Exchange and mirrored on the NYSE.

Does the Taiwan Stock Exchange require full disclosure of manufacturing issues in real time?

I had serious doubt's about Charlie's story on the 28nm shutdown in February (and really don't believe most of his stories) but this one seems to have legs.

If their was no shutdown then just say so. Not a 'no comment'.

 

[Dave Baumann]

Binning isn't the solution either, atleast not at this point. Yields aren't the best for the 7970 at this point and early stock shortages of the 7970 compared to the 7950 show this.

This is all fact is it??

Ever considered that 7970 is just the more popular SKU at the moment because people are gravitating to the fastest possible "new" thing?