NVIDIA Kepler speculation thread

At some (probably debatable) point, you simply need to take for granted a certain number of functioning brain cells. If a one letter difference is all it takes, then I could end up buying the 14 years old Quake II instead of Quake IV, right?

it's actually better than the usual modus operandi when the 50% slower card bore the same name as the faster one. two cards may even look physically identical but one has memory running at 333MHz and the other one has 800MHz memory. (let's say, a geforce GT220)

here's an horrendous rip off if you're unaware. of course, the fast model is the one benchmarked in launch reviews.
 
Well it's nothing we should be celebrating or encouraging, put it that way. I wonder if the tech press will even mention it as being an issue? They were giving AMD what-for over their naming for a while, especially with Barts and Cypress. I'm not quite sure how a huge performance gap between two different chips with a similar name - brought about by them being different segment chips - can be morally defensible.
 
Ofc, but i really doubt if clock was enough, Nvidia will have use a gk104 for the TI ... ( like the 670 in retail oc is faster of the 680 hard to say ofc )

Well, the GTX 660 Ti is also a way for NVIDIA to sell partially defective chips. So it increases "effective" yields, so to speak, generating revenue and filling a hole in their lineup—a hole that happens to be in a bit of a market sweet-spot.

Even if GK106 can get kind of close to it, it still makes plenty of sense from NV's point of view.
 
Well, the GTX 660 Ti is also a way for NVIDIA to sell partially defective chips. So it increases "effective" yields, so to speak, generating revenue and filling a hole in their lineup—a hole that happens to be in a bit of a market sweet-spot.

Even if GK106 can get kind of close to it, it still makes plenty of sense from NV's point of view.

Is really the memory controller is one of the bigger problem for defective chips ? the 660TI is " rumored " to have the same amount of SP of the 670.. just the memory controller have some part disabled.
 
Last edited by a moderator:
Is really the memory controller is one of the bigger problem for defective chips ? the 660TI is " rumored " to have the same amount of SP of the 670.. just the memory controller have some part disabled.

Well as you say, these are rumors, but it could also be that the chips can't run at 670 clocks, which I imagine must be pretty common.

Plus, memory controllers are tied to ROPs. Beyond that, it's hard to say without knowing how much redundancy there is and where.
 
Well as you say, these are rumors, but it could also be that the chips can't run at 670 clocks, which I imagine must be pretty common.

Plus, memory controllers are tied to ROPs. Beyond that, it's hard to say without knowing how much redundancy there is and where.
Rumors also say the 660 Ti will have the same clock speed as the 670 or slightly higher (980 MHz base).
 
Well as you say, these are rumors, but it could also be that the chips can't run at 670 clocks, which I imagine must be pretty common.

Plus, memory controllers are tied to ROPs. Beyond that, it's hard to say without knowing how much redundancy there is and where.

Reduncy on rop or memory controller defective ........ Personolly i m not a specialist of tsmc process.. but i can imagine if you want use a chips who dont meet the qualifcation, this is not your first pregorative .... you are more concerned about other parts..

I can be totally wrong, but memory controller ( 2x 32bits ), you cant disable any one of them, you need choose them ( due to L cache ), and as you said, they are tied to ROP, but you cant choose what ROP will be disabled alternatively.... but you can do it for meet Asic quality by disable alternatively any SMID on a sku .. ( defective and meet qualification or not is 2 things different , this is ofc linked first for the Asic quality and votlage, then come the problem of defective parts on a chip you can use with disabling some parts, but offtly this is the SMID who will be disabled for this use, not the memory controller or ROP ( there's no reason, ROP are down on some parts, or part of the memory controller, basically if the chip is good, the rop or memory controller, you will not have them who are defective .. and the minority of chip where it happend is so low, it is not financial viable to use them for make a product... I have never heard of a chip who had them defective at a point you want use it, but disable them and use what is working, ....

im sorry my english is not very good for explain what i have on my mind .. )
 
Last edited by a moderator:
Rumors also say the 660 Ti will have the same clock speed as the 670 or slightly higher (980 MHz base).

In my opinion, the combination of all those rumors (if they turn out to be true) do not exactly point in the direction of Nvidia having yield issues with GK104: same clocks, late market entry, only 1 mem-ctrl disabled.

TBH, that really sounds like very good strategy execution after the first confusion about 680 placement in the beginning of the year: Maximize GK104-wafers for higher pricing cards as long as possible, satisfy mobile demand (mostly) with GK107, which is small and has high no. of chips per wafer and delay the launch of GK106 as long as possible to just hit the OEM holiday cycle.
 
In my opinion, the combination of all those rumors (if they turn out to be true) do not exactly point in the direction of Nvidia having yield issues with GK104: same clocks, late market entry, only 1 mem-ctrl disabled.

TBH, that really sounds like very good strategy execution after the first confusion about 680 placement in the beginning of the year: Maximize GK104-wafers for higher pricing cards as long as possible, satisfy mobile demand (mostly) with GK107, which is small and has high no. of chips per wafer and delay the launch of GK106 as long as possible to just hit the OEM holiday cycle.

I agree, Nvidia has made a lot out of what should have been a very difficult time for them. Nobody should forget that when they are buying their 660 Ti, they are getting a far inferior card that what they would have got had Nvidia stuck to their original planned strategy and pricing, however.
 
jimbo75 said:
I agree, Nvidia has made a lot out of what should have been a very difficult time for them. Nobody should forget that when they are buying their 660 Ti, they are getting a far inferior card that what they would have got had Nvidia stuck to their original planned strategy and pricing, however.
I don't understand (and not because of gobbledygook that only hints at English ;)). They have a chip with equal game performance that's 20% smaller and that can do with a much cheaper board because only 2/3s of memory chips needed and lower power too (cooling definitely, power circuit probably.) In addition, they've always had a stronger brand marketing-wise.

How should this have been a difficult time? I thought it was obvious from the start that they had a killer opportunity here.
 
CarstenS said:
In my opinion, the combination of all those rumors (if they turn out to be true) do not exactly point in the direction of Nvidia having yield issues with GK104: same clocks, late market entry, only 1 mem-ctrl disabled.

TBH, that really sounds like very good strategy execution after the first confusion about 680 placement in the beginning of the year: Maximize GK104-wafers for higher pricing cards as long as possible, satisfy mobile demand (mostly) with GK107, which is small and has high no. of chips per wafer and delay the launch of GK106 as long as possible to just hit the OEM holiday cycle.
Yeah, if you can sell a small die at very high price and can't keep up with demand, there's no reason to hurry a lower margin equivalent. GK107 is an exception, of course, but absolutely needed for laptops, so that's a different case.

I missed the Nvidia quarterly results last Thursday (everybody here too apparently) and just got around reading the transcript:
http://seekingalpha.com/article/800...arnings-call-transcript?page=3&p=qanda&l=last
And we’re about to introduce and we just started ramping into it, our Kepler GPU that is targeted squarely at the gamers. And the segment starts out at about $99.00 and goes all the way up to about $249.00.

And those segments, we haven’t been able to address until now. And we’ve not been able to address it because we just haven’t had enough supply. And now that the supply is starting to catch up and gives us enough supply to launch into that segment and that segment is just dying for Kepler, and finally we’re be able to bring Kepler to them.
 
I don't understand (and not because of gobbledygook that only hints at English ;)). They have a chip with equal game performance that's 20% smaller and that can do with a much cheaper board because only 2/3s of memory chips needed and lower power too (cooling definitely, power circuit probably.) In addition, they've always had a stronger brand marketing-wise.

How should this have been a difficult time? I thought it was obvious from the start that they had a killer opportunity here.

This "killer opportunity" only existed because AMD is so utterly incompetent on various levels.

Tahiti should never have been so heavy on compute. What has AMD gained here? All that has happened is they've been forced to constantly drop prices.

6 months later they finally get good drivers for some of the most heavily benchmarked games. That's not good enough. Then their marketing dept comes up with the wonderful idea of naming a faster card the 7970 GHz Edition? That should have been the 7980. Why is this faster card (which absolutely destroys the 680 in compute) selling for less?

In the end, Nvidia got a "killer opportunity" out of the simple fact that Tahiti was too slow, had bad drivers for too long, is the wrong choice of chip anyway and AMD marketing is a complete joke. I also strongly suspect that they are gaining almost nothing out of being so far ahead in TTM, as people are waiting on Nvidia anyway. They might as well design larger, faster (pure gaming) chips and just plain beat Nvidia on performance because the current strategy has clearly failed.
 
http://www.xbitlabs.com/articles/graphics/display/geforce-gtx-660-ti_11.html#sect0

21% slower than the 670 with AA. Comes about right in-between the 7870 and 7950 boost (16% slower than the 7950 boost at 1080p AA).


http://hexus.net/tech/reviews/graphics/43701-powercolor-hd-7870-vortex-ii/?page=8 (this is not a 660 Ti benchmark)

The benchmark results are so heavily favoured toward AMD cards that NVIDIA actively encourages media not to use DiRT Showdown as a benchmark.
Anyone got a link to the most recent reviewers guide? I wonder if this is backfiring as quite a few sites seem to be actively benchmarking it now :p
 
Last edited by a moderator:
Back
Top